Boosting Bandwidth to Future-Proof CATV Solutions

Original blog post from:

Future entertainment systems and work-at-home environments are rapidly moving toward greater two-way interaction, which calls for enhanced downstream bandwidth and upstream capabilities. To stay competitive in the evolving CATV business, innovative technologies are needed to keep up with demands. One component that can play a vital role in this evolution is the CATV amplifier based on gallium nitride (GaN) technology. This post provides insight into how to do just that. The following is an excerpt from a Qorvo white paper, How to Increase Downstream Bandwidth and Upstream Capabilities in CATV Amplifiers with Greater Efficiency.

Meeting Higher Uplink Bandwidth Demands

Typical allocations for upstream traffic on Hybrid Fiber Coax (HFC) networks in the US range from 5 MHz to 42 MHz. User activities and new use cases are driving the need for increased capacity. In response, some multiple system operators (MSOs) are setting mid-splits or high-splits within the available bandwidth to accommodate these demands, reducing downstream bandwidth and possibly curtailing content or services.

MSOs facing this challenge are exploring the options available through DOCSIS 3.1 or DOCSIS 4.0 specifications. Upstream capacity within DOCSIS 3.1 can be extended up to 204MHz. While DOCSIS 4.0 allows the upstream to go up to 684MHz in both full duplex (FDX) and extended spectrum DOCSIS (ESD). Figure 1 shows how full duplex (FDX) let upstream and downstream traffic share the 684 MHz frequency range.

Take a Deeper Dive

Download the White Paper

Figure 1. DOCSIS 4.0 FDX spectrum for upstream (US) and downstream (DS).

Maintaining Linearity

The trend toward extended CATV bandwidth has led engineers and system architects to explore new technologies as networks are upgraded, requiring a newer generation of passive and active products. To meet demands for the increased bandwidth and data rates, CATV amplifiers must maintain a higher linear output power.

Gallium Nitride (GaN) devices can deliver more than the necessary efficiency and performance to satisfy DOCSIS requirements for CATV amplifiers. Consistent linearity is a primary requirement for reliable data transmission and signal integrity across HFC networks. The nonlinear behavior of active power devices can degrade the signal quality, leading to bit errors on digital channels and sometimes complete failure when trying to demodulate the signal.

Linearity of a gain block or amplifier depends primarily on these factors:

  • Semiconductor technology
  • Circuit design
  • Power consumption
  • Thermal design


A high degree of linearity and efficiency are paramount when designing an HFC amplifier, and this is where GaN-based components have a clear advantage. Figure 2 shows the fundamental components of a CATV amplifier. In terms of linear output, the downstream performance of an HFC amplifier or node largely relies on the output stage gain block (also called the power doubler).

What is DOCSIS?

Data Over Cable Service Interface Specification (DOCSIS) is an international telecommunications standard that permits the addition of high-bandwidth data transfer to an existing cable television (CATV) system. It is used by many cable television operators to provide Internet access over their existing hybrid fiber-coaxial (HFC) infrastructure. The version numbers are sometimes prefixed with simply “D” instead of “DOCSIS” (e.g., D3 for DOCSIS 3).

Figure 2. Block diagram of a CATV amplifier.

GaN represents an enabling technology for power amplifier designs that can accommodate the demands of the DOCSIS 3.1 and DOCSIS 4.0 standards. Figure 3 illustrates a design stage that originally implemented the gain block architecture using field-effect transistors (FETs) based on GaAs technology. Replacing FET3 and FET4 with GaN-based components results in a substantial performance improvement because of the characteristics of these devices, including operation at high frequencies, high-voltage ruggedness, high current density, and power handling. GaN supports up to 10 W/mm compared to 1 W/mm for GaAs.

Figure 3. Gain block architecture improved through the use of GaN-based FETs.

Comparing the relative characteristics of material technologies used in CATV gain-block architecture, Figure 4 shows the advantages of GaN technology, enabling MSOs to boost linear output while maintaining existing amplifier spacing. This can minimize upgrade costs and make it possible to implement fiber deep solutions, locating fiber closer to the customer to enhance service and at the same time reducing or eliminating amplifiers.

Figure 4. Characteristics of GaN-based components compared to other options.

Qorvo Expertise in CATV Gain Blocks

Two Qorvo products are outstanding design choices in this sector:

  • QPA3260 power doubler hybrid – Delivers the highest linear output up to 1.2 GHz
  • QPA3315 power doubler hybrid – Supports DOCSIS 4.0 implementations in designs requiring up to 1.8 GHz capabilities


With the need for greater downstream bandwidth and increased upstream capacities spurred by DOCSIS 3.1 and DOCSIS4.0 standards, Qorvo can help MSOs respond to the challenge with GaN-based gain blocks for CATV signal amplification. CATV equipment manufacturers can rely on proven solutions that deliver better data transport using durable and reliable components.


Have another topic that you would like Qorvo experts to cover? Email your suggestions to the Qorvo Blog team and it could be featured in an upcoming post. Please include your contact information in the body of the email.

Explore the Qorvo Blog

About the Author

Rainer Hillermeier
General Manager, Design and Operations – Qorvo Germany

Rainer manages Qorvo’s Design and Manufacturing Site in Nuremberg, Germany. With more than two decades of high power CATV gain blocks design experience, Rainer oversaw the development and release of the first 1 GHz, 1.2 GHz and now 1.8 GHz gain blocks. He also pioneered the introduction of GaN semiconductor technology in hybrid fiber-coax infrastructure products.

Lights, Camera, Action! UWB is the Star Supporter in Le Premier Royaume

With all the potential use cases for ultra-wideband (UWB) technology, it’s not surprising that it’s often found in the most unexpected places. Case-in-point: the Puy du Fou theme park in France. A recent article in blooloop outlines how UWB rescued its show “Le Premier Royaume” (“The First Kingdom”), a multi-sensory walkthrough experience set in fifth-century France, in which guests journey through the history and legend of the Frankish King Clovis.

According to the article, “During the experience, five hundred theater goers move through the sets, as eight actors execute perfectly timed entrances and exits. This is supported by lights, sound, and special effects.” The show cycles for seven hours a day, using 14 different sets throughout roughly 24,000 square feet on two levels. The challenge for the show producers was to keep pace with the visitors moving through the spaces and timing the activation of the show’s effects.

That’s where UWB came in. Puy du Fou turned to Eliko’s UWB RTLS system, which integrates Qorvo’s UWB devices and can track hundreds of objects in real-time.

Here we explore why UWB technology is conducive for precision-dependent use cases.

figure 1

Eliko’s UWB RTLS system integrates Qorvo’s UWB devices and can track hundreds of objects in real-time.

UWB: Solving the ‘Where’

In show business, timing is everything. Especially for the Le Premier Royaume performance. What’s also critical for the show’s success is the ‘where.’ The timing of the lighting, special effects, actors are all based on where the audience is located in relation to its location within the theater. We have many technologies—BLE, Wi-Fi, RFID, etc.—and apps built into so many of our devices that accurately track the ‘when’ and inform us of ‘what.’ But they’re not designed to track the ‘where,’ the precise location in real-time—second by second. However, this is exactly what ultra-wideband does.

With UWB, the show’s production could synch all the devices so the apps always know precisely where things are—down to the centimeter and the second. Being a low-power, ultra-wide bandwidth radio technology, UWB offers characteristics that make it ideal for use cases such as Puy du Fou’s—the most important in this case are precision ranging, precise location, and fast data communication.

Why UWB?

To better understand why UWB, in particular, is the most conducive for applications like Puy du Fou’s – a challenging environment for RF technologies – let’s take a look at how it compares to other technologies like BLE and Wi-Fi. This table shows that UWB has all the ingredients to overcome the show’s challenges.

figure 1

A scene from “Le Premier Royaume” (“The First Kingdom”), a multi-sensory walkthrough experience set in fifth-century France.

UWB Bluetooth® Low Energy Wi-Fi
Accuracy Centimeter 1-5 meters 5-15 meters
Reliability Strong immunity to multi-path and interference Very sensitive to multi-path, obstruction and interference Very sensitive to multi-path, obstruction and interference
Range/coverage Typ. 70m
Max 250m
Typ. 15m
Max 100m
Typ. 50m
Max 150m
Data Communications Up to 27 Mbps Up to 2 Mbps Up to 1 Gbps

What sets UWB apart the most from the other technologies is its accuracy. Like other location technologies, UWB doesn’t rely on satellites for communication. Instead, devices containing UWB technology (like antennas on the show’s actors, lighting, cameras, etc.) communicate directly with each other to determine location and distance. What’s different about UWB from the others is this accuracy is achieved by measuring the time that it takes signal pulses to travel between devices, which can be calculated based on the time-of-flight of each transmitted pulse.

The accuracy of this method depends on the signal’s bandwidth; very wide signals are needed to achieve high accuracy. UWB signals use roughly 500 MHz of bandwidth—many times wider than other technologies sometimes used for location sensing. That enables UWB to achieve centimeter precision, which is critical for many applications.

UWB technology also works in non-line-of-sight conditions where the use of cameras to track location is not possible. This means that the signal is capable of going through obstructions like the sets while maintaining a very high location accuracy. In addition, as it operates at frequencies between 6 and 8GHz, it has no interference issues with other radio waves.

For more in-depth information about how UWB technology works, read the whitepaper, Getting Back to Basics with Ultra-Wideband (UWB).

Check Out Other UWB Use Cases

UWB is an old radio technology enabling new opportunities for rich real-time applications. Companies and application developers are already enabling new UWB-based services that benefit both businesses and individuals. But they are only scratching the surface. There are many use cases where UWB can create an experience and enrich our personal lives. For more ideas on use cases, watch this video. You can also read about how UWB is used at the Museum of the Bible to enhance the visitor experience, as well as how Volkswagen is optimizing its production by integrating UWB into its operations.

The Bluetooth® word mark and logos are registered trademarks owned by Bluetooth SIG, Inc. and any use of such marks by Qorvo US, Inc. is under license. Other trademarks and trade names are those of their respective owners.

What Designers Need to Know to Achieve Wi-Fi Tri-Band Gigabit Speeds and High Throughput

Original post from:

Engineers are always looking for the simplest solution to complex system design challenges. Look no further for answers in the U-NII 1-8, 5 and 6 GHz realm. Here we review how state-of-the-art bandBoost™ filters help increase system design capacity and throughput, offering engineers an easy, flexible solution to their complex designs, while at the same time helping to meet those tough final product compliance requirements.

A Summary of Where We are Today in Wi-Fi

Wi-Fi usage has grown exponentially over the years. Most recently, it has skyrocketed upward to unimaginable levels — driven by the pandemic of 2020 due to work from home, school requirements, gaming advancements, and, of course, 5G. According to Statista, the first weeks of March 2020 saw an 18 percent increase in in-home data usage compared to the same period in 2019, with average daily data usage rates exceeding 16.6 GB.

With this increase in usage comes an increase in expectations to access Wi-Fi anywhere — throughout the home, both inside and out, and at work. Meeting these expectations requires more wireless backhaul equipment to transport data between the internet and subnetworks. It also requires advancements in existing technology to reach the capacity, range, signal reliability and the rising number of new applications wireless service providers are seeing. Figure 1 shows the exponential increase in wireless applications — from email to videoconferencing, smart home capabilities, gaming and virtual reality — as wireless technology continues to advance.

Go In Depth:



Figure 1: The advancement of Wi-Fi

The 802.11 standard has now advanced onto Wi-Fi 6 and Wi-Fi 6E, providing service beyond 5 GHz and into the 6 GHz area up to 7125 GHz, as shown in Figure 2. This higher frequency range increases our video capacities for our security systems and streaming.

Figure 2: Tri-Band Wi-Fi frequency bands

However, working in higher frequency ranges can bring challenges such as more signal attenuation and thermal increases — especially when trying to meet the requirements of small form factors. To meet these challenges head-on, RF front-end (RFFE) engineers need to take existing technology to another level. One of those advancements has been in BAW filter technology now being used heavily in Wi-Fi system designs.

As shown in Figure 3 below, Qorvo has three BAW filter variants that boost overall Wi-Fi performance, maximize network capacity, increase RF range, and mitigate interference between the many different in-home radios operating simultaneously.

Figure 3: bandBoost, edgeBoost, and coexBoost filter technology performance

5 & 6 GHz bandBoost Filters

In a previous blog post called An Essential Part of The Wi-Fi Tri-Band System – 5.2 GHz RF Filters, we explored how using bandBoost filters like the Qorvo QPQ1903 and QPQ1904 can help reduce design complexity and help with coexistence. We also explored how these bandBoost filters provide high isolation, helping to reduce that function on the antenna design, allowing for less expensive antennas. Therefore, the RFFE isolation parameter no longer needs to rest entirely on the antenna. This reduces antenna and shielding costs – providing up to a 20 percent cost reduction.

These bandBoost BAW filters play a key role in separating the U-NII-2A band from the U-NII-2C band, which only has a bandgap of 120 MHz, as shown in Figure 4. Using these filters, we can attain Wi-Fi coverage reaching every corner of the home with the highest throughput and capacity. Using this solution in a Wi-Fi system design has shown increases in capacity for the end user up to 4-times.

Unlicensed National Information Infrastructure (U-NII)

The U-NII radio band, as defined by the United States Federal Communications Commission, is part of the radio frequency spectrum used by WLAN devices and by many wireless ISPs.

As of March 2021, U-NII consists of eight ranges. U-NII 1 through 4 are for 5 GHz WLAN (802.11a and newer), and 5 through 8 are for 6 GHz WLAN (802.11ax) use. U-NII 2 is further divided into three subsections: A, B and C.

Figure 4: 5 GHz bandBoost filters and U-NII 1-4 bands

These filters are much smaller than legacy filters on the market used in Wi-Fi applications — allowing for more compact tri-band radios. They also have superior isolation achieving greater than 80 dBm system isolation for designers. This helps engineers meet the stringent Wi-Fi 6 and 6E requirements.

Figure 5: Benefits of using QPQ1903 and QPQ1904 bandBoost filters

The addition of multiple-input multiple-output (MIMO) and higher frequencies in the 6 GHz range increases system temperatures. With more thermal requirements, robust RFFE components are a must. Much of the industry specifies their parts in the 60°C to 80°C range, but higher temperature operation is needed based on the system temperatures produced in this frequency range. To solve these challenges, many hours of design effort have been spent on increasing the temperature capabilities of BAW. As product designs in Wi-Fi 5, 6/6E, and soon to come Wi-Fi 7, development has become more challenging, and as new opportunities like the automotive area opened for BAW, the push for higher temperature capability has come to the forefront.

Qorvo BAW technology engineers have delivered innovative devices by designing those that exceed the usual 85°C maximum temperature working range, going up to +95°C. The benefits this creates are great for both product designers and end-product customers. Now sleeker devices are achievable, as end-products no longer require large heat sinks. This also reduces design time as engineers can more easily attain system thermal requirements. One other advancement related to heat is that the bandBoost BAW products work at +95°C while still meeting a 0.5 to 1 dBm insertion loss.

This lower insertion loss improves Wi-Fi range and receive quality by up to 22 percent. Lower insertion loss also means improved thermal capability and performance as the RF signal seen at the RFFE Low Noise Amplifier (LNA) is improved. Below, Figure 6 shows the features and benefits of the QPQ1903 and QPQ1904 edgeBoost™ BAW filter.

Figure 6: Features and benefits of QPQ1903 and QPQ1904

Not only are these filters providing benefits to the LNA, but they are small and perform well enough to install inside a tiny integrated Wi-Fi module package housing the LNA, switch, PA, and filter. Doing this drastically changes the end-product system layout making design easier and helps speed time-to-market. No longer are engineers burdened with matching and plugging individual passive and active components onto their PC board, but now they have all that done in these complex integrated modules called integrated front-end modules (iFEMs), creating a plug-and-play solution easily installed on their design.

A perfect example of this is the QPF7219 2.4 GHz iFEM, as seen in Figure 7. Qorvo has not only provided solutions with discrete edgeBoost BAW filters to increase output and capacity across all Wi-Fi channels. But Qorvo has gone a step further by including this filter inside an iFEM, our QPF7219, to provide customers with a drop-in pin-compatible replacement providing the same capacity and range performance outcome. This provides customers with design flexibility, board space in their design and is the first one of its kind on the market.

Figure 7: edgeBoost used as discrete and inside an iFEM

The need for smaller and sleeker product designs is always top of mind for Wi-Fi engineers. But to achieve the goal means component designers need to develop smaller products in many areas of the design, not just in one or two areas. From a tri-band Wi-Fi chip-set perspective, Qorvo has addressed this head-on. Qorvo has provided an entire group of iFEM alternatives to address the many signal transmit and receive lines in a product. This allows Wi-Fi design manufacturers to manage all the UNII and 2.4 GHz bands in a tri-band end-product design.

Figure 8: 2.4 & 5 GHz Wi-Fi 6 with IoT Tri-Band front-end solutions

This new design solution of combining the filter inside the iFEM equates to a smaller PC board and less shielding, as shown in Figure 9 below. Shielding matching and PC board space are expensive, not to mention the additional time associated with providing these materials. By placing all the RFFE materials inside a module, system designers can save cost, design faster, and get their products to market more quickly.

Figure 9: Putting the filter technology inside the iFEM removes shielding and reduces overall RFFE form-factor

As Wi-Fi system designers continue to be challenged with new specification requirements, they need newer or enhanced technologies to meet the need. By collaborating with our customers, we have provided state-of-the-art solutions to solve the tough thermal, performance, size, interference, capacity, throughput, and range difficulties seen by their end-customers. These solutions enable them to improve their designs to meet the Wi-Fi wave of today and in the future.


About the Author

Igor Lalicevic
Senior Marketing Manager, Wireless Connectivity Business Unit

With over 20 years of experience in the wireless industry, Igor helps Qorvo engineering teams create state-of-the-art RF components and solutions. He inspires the creation of new wireless connectivity products and eco-systems innovations that make a deep impact on our everyday life.

Why GaN is 5G’s Super ‘Power’

While some feel GaN is still a relatively new technology, many can’t dispute how it’s advanced to the head of the class. AKA, Gallium Nitride, GaN is a technology on the cusp of dethroning silicon LDMOS, which has been the material of choice in high power applications. GaN is a direct bandgap semiconductor technology belonging to the III-V group. It is increasingly being used in power electronics because of its higher efficiency, superior high-voltage sustainability, reduced power consumption, higher temperature attributes, and power-handling characteristics.

These attributes have thrust GaN into the 5G RF spotlight – especially when it comes to mmWave 5G networks. And, while we all have ‘heard’ the promises of 5G, today, many of us in big cities – about 5 million of us to be more precise – are starting to realize those promises as major wireless carriers roll 5G out to their customers. But we are not there yet. Not even close. The goal is to connect 2.8 billion users by 2025. To reach this goal means to revamp the entire mobile infrastructure – a complex undertaking. But it can be done. And with the help of GaN technology, 5G will be in billions of people’s hands before you know it.

Recently, invited Qorvo’s own Roger Hall to pen a series of 5G articles that explain the complexities of building out the infrastructure and where GaN fits into the innovations that will bring 5G to the masses. Here are summaries of each article with a link for a deeper dive.

5G and GaN: Understanding Sub-6 GHz Massive MIMO Infrastructure

In this article, Roger explains the advantages for carriers to implement Massive MIMO technology as a means to minimize cost and increase capacity when rolling out 5G. He explores sub-6 GHz and why it’s important for increasing the adoption and expansion of 5G. He also addresses how GaN is being used in Massive MIMO Infrastructure applications. Read more >

5G and GaN: The Shift from LDMOS to GaN

Here Roger examines how the power demands of sub-6 GHz 5G base stations are driving a shift from silicon LDMOS amplifiers to GaN-based solutions, and what makes GaN a viable technology for many RF applications. Roger also reviews some of the tradeoffs engineers need to consider between these two technologies and why GaN is becoming the clear winner in many 5G solutions. Read more >

5G and GaN: What Embedded Designers Need to Know

Building on the previous article, Roger provides insight for embedded designers to fully realize the potential of GaN. He discusses misconceptions about GaN, explores its characteristics, and offers best practices to maximize its performance. Read more >

5G and GaN: Future Innovations

In his fourth and final article in this series, Roger looks to the future of GaN’s role in base stations. He provides a peek into GaN innovations being made today that will improve linear efficiency, power density and reliability and the implications of those improvements. Read more >

For more information on GaN technology, visit here.

About the Author

About Roger Hall
Roger Hall

Roger is the General Manager of High-Performance Solutions at Qorvo. He leads program management and applications engineering for Wireless Infrastructure, Defense and Aerospace, and Power Management markets. This overarching role gives him a unique lens to view and interpret where RF technologies play fundamental parts in enabling future innovations.

Qorvo Blog Team

One part technical, one part content, and one part strategic, our small team is dedicated to connecting you with helpful, timely insights from some of the bright minds at Qorvo.

New capacitive sensing technology provides a more reliable and safer hands-on detection

The AS8579 sensor offers the simplest way for car makers to comply with the UN Regulation 79, while giving the best detection performance

For automotive design engineers, it is unusual to find a new technology solution which performs better than existing approaches, and which reduces cost, and which is easier to implement in the application. But that is exactly what a new capacitive sensing chip, ams’ new AS8579, offers when used for hands-on detection (HOD) in cars which provide driver assistance functions.
It is the result of the application of a familiar and proven measurement principle – I/Q demodulation – to the job of sensing the position of the driver’s hands on the steering wheel. And it is markedly superior to any of the existing technologies in use for HOD in cars. Watch the highlights in our video:


Essential safety requirement in new car designs

The HOD function is required by the United Nations Regulation 79, and applies to all new cars that have a Lane Keeping Assist System (LKAS) wherever ratified. It has already been adopted by the European Union for new production vehicles from 1 April 2021. The purpose of the HOD system is to continuously monitor the readiness of the driver to assume control of the steering system in an emergency, or in the event of the failure of the LKAS.

Various technologies have been developed to provide this HOD function, but have had limitations: it is possible for drivers who want to avoid holding the steering wheel to fool the current monitoring system, which could compromise safety. And some existing solutions also perform poorly in certain operating conditions.

One approach to HOD has been the torque sensor: this detects the continual, minute deflections produced when the driver grips the steering wheel. The big drawback of this technology is that it can be easily fooled: the driver may take their hands off the wheel and ‘hold’ it by pressing upwards against it with their leg.

The problems with torque sensors have led the car industry to adopt a form of capacitive sensing for HOD: it monitors the driver’s grip on the steering wheel by detecting the change in capacitance of the steering wheel when the driver’s hands – which absorb electrical charge – come into contact with it. This technique only requires a single sensor chip connected to a metal sensor element built into the steering wheel.

Until now, automotive system manufacturers have used the charge-discharge method of capacitive sensing: this is a well understood technique, as it has been applied for many years in products such as touchscreens and touch-sensing buttons. But detection fails when the driver wears gloves, and false detection signals generated by the presence of moisture or humidity undermine the safety performance of hands-on detection based on this method of capacitive sensing. This type of capacitive sensor can even be fooled if the driver wedges a capacitive object, such as a piece of fruit or a plastic water bottle, into the frame of the steering wheel. So again, the implementation of this charge-discharge method of capacitive sensing potentially compromises safety.

It is true that other technologies are already applied to other driver-monitoring functions. For instance, 2D optical sensing is in use in systems for monitoring the position of the driver’s head. However, these 2D optical-sensing systems are not capable of performing HOD. This means that capacitive sensing is the most viable technology for HOD that is ready for deployment today. And now ams has a new approach to capacitive sensing which will meet all the safety requirements imposed by the automotive industry, and which is simple to implement.

Better performance, lower cost

This new solution from ams provides better performance, and with fewer components than the existing charge-discharge technique for capacitive sensing.

By implementing reliable capacitive sensing based on I/Q demodulation, the AS8579 capacitive sensor performs HOD in a way which cannot be fooled. Like the charge-discharge method, I/Q demodulation is a proven and well-known technique for capacitive sensing. Its advantage is that it measures the resistive as well as the capacitive element of a system’s impedance. The effect of this is that, unlike the charge-discharge method, it works reliably in difficult conditions, such as in the presence of moisture, or when the driver is wearing gloves. And it cannot be fooled, so provides for assured detection of the driver’s grip on the steering wheel. And the added benefit of the AS8579-based solution is that it can operate via a heated steering wheel’s heater element, so it does not require a separate sensor element to be built into the steering wheel.

This is how the AS8579 eliminates the normal trade-offs in engineering design:

  • It performs better – it cannot be fooled, and it operates in all conditions
  • It costs less – it is a single-chip solution, and requires no dedicated sensing element in a heated steering wheel
  • It is easy to implement – the chip’s output is an impedance measurement, and the system controller simply applies a threshold value to determine whether hands are on the steering wheel or not.

Ready for use in automotive designs

The AS8579 is fully automotive qualified, and offers multiple on-chip diagnostic functions, ensuring support for the ISO 26262 functional safety standard up to ASIL Grade B. Operating at one of four selectable driver-output frequencies – 45.45kHz, 71.43kHz, 100kHz or 125kHz – the AS8579 offers high immunity to electromagnetic interference.

Automotive designers can start developing with the AS8579 automotive capacitive sensor immediately using its dedicated evaluation kit, the AS8579-TS_EK_DB. 

For more technical information or for sample requests, please go to capacitive sensor AS8579.



1D ToF family for mobile and industry brings the right combination of performance, size, and cost

Original blog post:

ams 1D time-of-flight ranging sensor family offers mobile and industrial customers the right combination of performance, size, and cost to meet their needs.

When innovating sensor technology for a better lifestyle, ams engineers are balancing three attributes that are vital to customers: sensor performance, sensor size, and system cost. These variables are almost infinitely adjustable according to our customers’ evolving needs and specifications, competitive conditions, regulatory constraints or bill of materials requirements. How we design our product portfolio is based on our reading of the market and what our customers and close design partners tell us they want.

Customers are in a never-ending race to deliver better products and experiences. And as part of this, 1D Time-of-Flight sensors for front and world-facing applications are becoming increasingly important in the mobile, consumer, wearables, PC and industrial segments. The ams family of 1D ToF ranging sensors, developed for Laser Distance Auto-Focus (LDAF) applications within the mobile phone industry area also bringing benefit and increasingly winning in applications including PC user detection enabling auto lock/unlock, obstacle avoidance in robotic vacuum cleaners, inventory management, to name a few.

Broadening the family of time-of-flight (ToF) ranging sensors

ams has a strong history in bringing 1D ToF sensing innovations to market. Our most recent innovations include the world’s smallest 1D ToF sensor for accurate proximity sensing and distance measurement in smartphones’ – the TMF8701 and the TMF8801 which extends the operating range of the direct time of flight module to enable smartphones with space-saving accurate distance measurement. Now, ams brings rounds out the TMF sensor family with the TMF8805 adjusting the performance/cost variables to give customers greater flexibility and choice, especially for applications and products with massive growth potential or competing in the uncertainty of emerging markets.

TMF8805 – for mobile phone camera applications and more

The TMF8805 is a highly-integrated module which includes a class 1 eye-safe 940nm Vertical Cavity Surface Emitting Laser (VCSEL), Single Photon Avalanche Diode (SPAD) array, time-to-digital converter (TDC) along with a low power, high performance microcontroller. This system-in-module integration enables robust and precise distance measurements in the 20mm and 2500mm range, all packaged in the industry’s smallest footprint measuring only 2.2mm x 3.6mm x 1.0mm.

This high precision distance measurement is ideal for use in world-facing, LDAF mobile phone applications by enabling a fast, high-precision auto-focus feature. The new sensor joins the existing TMF8801 and TMF8701 time-of-flight sensors from ams, providing products which meet a range of cost and performance requirements across the mobile, wearable and consumer electronics, computing and industrial markets.

To meet evolving expectations in a transforming world, customers come to ams for our simple-to-integrate, plug-and-play sophisticated sensor systems, while often benefiting from the ‘speed premium’ of our supplier ecosystem and specialist expertise. The TMF8805 time-of-flight sensor is now in mass production and an evaluation kit featuring the TMF8805 along with a comprehensive evaluation GUI is also available.


The Challenges 5G Brings to RF Filtering

Blog post from:

In the race to implement mainstream 5G wireless communication, the world is waiting to see if this next-generation network will achieve a hundredfold increase in user data rates. This transformative technology not only boosts performance for the latest cell phones, but also for fixed wireless access (FWA) networks and Internet of Things (IoT) smart devices. In order to reach 10 Gbps peak data rates, the increase in channel capacity must come from somewhere. A key innovation at the heart of 5G is utilizing new frequencies greater than 20 GHz in the millimeter wave (mmWave) spectrum, which offers the most dramatic increase in available bandwidth.

A well-known downside to high frequencies is the range limitation and path loss that occurs through air, objects, and buildings. The key workaround for mmWave base station systems is the use of multi-element beamforming antenna arrays in both urban and suburban environments. Since mmWave signals require much smaller antennas, they can be tightly packed together to create a single, narrowly focused beam for point-to-point communication with greater reach.


In order to overcome the range limitations of mmWave frequencies, dense arrays of antennas are used to create a tightly focused beam for point-to-point communication.


Changes in Radio Architecture for mmWave Beamforming

Of course, these new beamforming radio architectures bring a whole new set of challenges for designers. Traditional filtering solutions for radios operating in the ranges of 700 MHz and 2.6 GHz are no longer suitable for mmWave frequencies such as 28 GHz. For example, cavity filters are often used in the RF front ends of LTE macrocells. However, an inter-element spacing of less than half the wavelength (λ/2) is required to avoid the generation of grating lobes in antenna array systems. This inter-element spacing is about 21.4 cm at 700 MHz frequency versus only about 5 mm at 28 GHz. Such a requirement therefore calls for very small form factor components in the array. Not only will the whole RF front end be reduced in size, but also the number of RF paths will be increased – which means the filters right next to the antennas must be very compact.

Screen Shot 2019-09-26 at 11.35.39 AMAs the frequency increases, the antenna size must decrease, leading to significant changes in the mmWave beamforming radio architecture.


Addressing Challenges with mmWave Filtering

Based on decades of experience working with mmWave filtering solutions, Knowles Precision Devices has a product line of mmWave filters solutions that addresses these challenges. Using specialized topologies and material formulations, we’ve created off-the-shelf catalog designs available up to 42 GHz that are 20 times smaller than the current alternatives.

Screen Shot 2019-09-26 at 11.44.40 AM

Compared to current alternatives, Knowles Precision Devices offers filter solutions that are 20 times smaller to meet mmWave inter-element spacing requirements.


As manufacturers push to increase available bandwidth, temperature stability also becomes more and more of an issue. MmWave antenna arrays may be deployed in exposed environments with extreme temperatures, and heat dissipation issues may arise from packing miniature components onto densely populated boards. In order to guarantee consistent performance, our filters are rated for stable operation from -55°C to +125°C. For example, the bandpass filters below shifted only by 140 MHz when tested over that temperature range.

Screen Shot 2019-09-26 at 11.30.04 AM

The temperature response of a Knowles 18 GHz band pass filter (BPF) on CF dielectric shows little variation in performance from -55°C to +125°C.


Finally, high performance with high repeatability is key to ensuring the best spectral efficiency and rejection possible. High frequency circuits are especially sensitive to variations in performance, so precise manufacturing techniques ensure that our filter specifications – such as 3 GHz bandwidth and greater than 50 dB rejection – are properly maintained from part to part. Plus, unlike chip and wire filter solutions, our surface mount filters standardize the form factor, reduce overall assembly time, and do not require tuning – thus saving in overall labor costs and lead times.


The performance of Knowles Precision Devices 37-40 GHz BPFs shows highly repeatable performance over 100 samples, even without post-assembly tuning.


An Essential Part of The Wi-Fi Tri-Band System – 5.2 GHz RF Filters

Blog post from our partner: Qorvo


As engineers, we’re always looking for the simplest solution to complex system design challenges. Look no further for solutions in the 5.2 GHz Wi-Fi realm. Here we step you through the resolutions to reduce design complexity, while meeting those tough final product compliance requirements.

An Essential Part of The Wi-Fi Tri-Band System
– 5.2 GHz RF Filters

Most individuals using Wi-Fi in any setting – home, office, or coffee shop – expect a fast upload and download experience. To achieve this, individuals and businesses must move toward a tri-band radio product solution. Anyone familiar with the two Wi-Fi bands of 2.4 GHz and 5 GHz, know the higher 5 GHz band is faster for upload and download speeds. However, the use of higher frequency bands comes at a price of signal attenuation. They also have a high probability of interference with other closely aligned spectrum.

This is where Qorvo filter technology comes in. Qorvo filters help mitigate these probable interferences. In this blog, we focus on the 5.2 GHz filter technology, the higher area of the Wi-Fi tri-band. This filter technology improves the mesh network quality-of-service and helps meet system regulatory requirements.

Wi-Fi and Cellular Frequency Bands

The 5 GHz band provides over six times more bandwidth than 2.4 GHz. This is a big plus for today’s video streaming, chatting, and gaming, which is in high demand. As shown in the figure below, spectrum re-farming and the crowding of wireless standards has increased. As Wi-Fi evolves into even higher realms of 5 GHz and 6 GHz, the crowding continues. As seen in the below figure, the 5 GHz unlicensed area clearly must coexist with cellular frequencies on all sides. One such area is the 5.2 GHz band.

Wireless Standards Frequency Band Infographic
Figure 1: Major wireless frequency bands for 5G Frequency Range (FR) 1 & 2

The Wi-Fi 5.2 GHz Space

The Wi-Fi 5 GHz UNII (Unlicensed National Information Infrastructure) arena is primarily used in routers in retail, service providers and enterprise spaces. These routers are commonly found in mesh networks, extenders, gateways and outdoor access points. In the UNII1-2a band (i.e. 5150-5350 MHz – 5.2 GHz band), maintaining a minimum RF system pathloss is important. Providing a minimum filter insertion loss is imperative to reduce power consumption, achieve better signal reception, and decrease thermal stress. This helps system designers to deliver low carbon footprint end-products. Additionally, for the 5.2 GHz band, a high out-of-band filter rejection is desired – notably around 50 dB or greater at 5490-5850 MHz – which helps mitigate crosstalk and enables coexistence with the 5.6 GHz UNII band, as shown in Figure 2 below.

5.2 GHz Filter Deep-Dive Analysis (QPQ1903)

To meet the need of the true mesh tri-band application, a well-designed filter is required. Today, few filter suppliers can meet the standard’s body specifications of rejection, insertion loss and power handling in the 5.2 GHz band. As shown in the below figure, Qorvo provides several filter solutions for the UNII 5 GHz bands.

5 GHz RF Filters Infographic
Figure 2: 5 GHz UNII frequency bands, bandwidths, bandedge parameters, and filters

A Review of The Wi-Fi Standard Specifications

High Rejection – A high rejection is critical in a system for two main reasons – one being the need to mitigate interference, and two, to mitigate unwanted signal noise. Therefore, customers and standards bodies have set a desired high parameter of 50 dB or greater on rejection rate for critical out-of-band signals. But it is important to achieve this without losing the integrity of the RF signal range and capacity. Using Qorvo bandBoost filters such as the QPQ1903 achieves this goal.

Learn More about Qorvo’s bandBoost filters:

Download the Infographic

As noted in Figure 2 above, coexistence with the Wi-Fi 5.6 GHz band has only a 120 MHz gap. Meeting the 50 dB or greater rejection at this frequency (5490-5850 MHz) within this small gap is challenging. It requires filter technologies with steep out-of-band skirts and a high rejection rate. BAW (bulk acoustic wave) SMR (solidly mounted resonator) technology is well suited to meet thermal and high rejection filter requirements. BAW-SMR has a vertical heat flux acoustic reflector that allows thermal heat to dissipate away quickly and efficiently from the filter. BAW exhibits very little frequency shifts due to self-heating as the topology of BAW-SMR provides a lower resistance, preventing the resonators from overheating. As shown in Figure 3 below, the rejection rate of the 120 MHz gap is greater than 50 dB, well within the regulatory guidance parameters.

Graph of the QPQ1903 5.2 GHz Filter's Electrical Performance
Figure 3: 5.2 GHz QPQ1903 performance measurements @ 25⁰C
Insertion Loss – In Figure 3 above, the insertion loss performance data for the 5.2 GHz bandBoost™ filter, QPQ1903, is 1.5 dB or better. Thus, providing improved system pathloss and power consumption. This translates directly to increased coverage range, better reception of the signal and simplified thermal management of the end product. Additionally, the return loss (not shown) meets the required specification by 2 dB or better across the entire pass band. This allows for an easier system match when using a discrete 5.2 GHz filter solution, which is especially critical in high linearity Wi-Fi 6 systems.

Power Handling – Power handling has become a major requirement for today’s wireless systems. 5G and the increases in RF input power levels to the receiver have attributed heavily to this new high-power handling system requirement. Thus, RF filters must be able to meet higher input power levels, sometimes up to 33 dBm. Luckily due to BAW-SMR’s ability to dissipate heat efficiently, it can easily attain this goal without compromising performance.

Not only has Qorvo BAW technology met the required critical regulatory specifications, but in most cases has done so with additional margin, and in the higher temperature ranges of up to 95⁰C – further helping customers and system designers meet stringent final product thermal management requirements.

New Wi-Fi System Complexity Challenges

The onset of Wi-Fi 6 and 6E standards has introduced additional application and system design complexity. Some of the most common challenges are:

  • Tri-band Wi-Fi
  • Smaller, sleeker device form factors
  • Higher temperature conditions due to form factor and application area requirements
  • Completely integrated solutions with no external tuning

Tri-band Wi-Fi – The need for data capacity and coverage has led to an explosion of the tri-band architecture of 2.4 GHz, 5.2 GHz and 5.6 GHz bands in gateways and end-nodes, as shown in Figure 4 below. It allows users to connect more devices to the internet using the 5 GHz spectrum in a more efficient way. For example, if you’re using a home mesh system with multiple routers to cover a larger space, the second 5.6 GHz band acts as a dedicated communications line between the two routers to speed up the entire system by as much as 180% over dual-band configurations.

5 GHz Wi-Fi 6 Spectrum Filters Response Graph
Figure 4: 5.2 GHz & 5.6 GHz Wi-Fi filter response bandwidths and bandedge parameters
Higher temperature conditions – New gateway designs require smaller form factors, sometimes half the size of previous product versions, with a higher number of RF antenna pathways (a tri-band router can have upwards of 8 RF pathways). Because of the confined space and increased number of antenna pathways, system thermal management becomes even more challenging. There are also gateway applications with higher temperatures at the extremes of -20⁰C to +95⁰C for outside environments.

Smaller form factor & integration – Today’s gateway devices are sleeker, stylish and have smaller form factors. The demand for smaller gateway form factors is pushing Wi-Fi integrated circuits to shrink as well. Additionally, this is moving semiconductor technologies toward creating smaller and thinner devices to mitigate system designer costs.

To accommodate this initiative, Qorvo has also integrated its filter technology into complete RF front-end designs to help system designers reduce design time, meet smaller size requirements, and increase time-to-market.

Comparison of 5 GHz ceramic versus Qorvo QPQ1903 BAW filters
Figure 5: Comparison of 5 GHz ceramic versus Qorvo QPQ1903 BAW filters
As system designers continue to meet higher frequency demand initiatives – brought on by the hunger for more frequency spectrum – design challenges increase. New hurdles like system pathloss, signal attenuation, designs in higher frequency realms and meeting coexistence standards are becoming more prevalent. Qorvo has worked extensively with network standards bodies and customers across the globe to ensure their designs meet all compliance needs. We work hard to provide products that meet these highly sought-after specifications with additional margin. This proactive work ultimately allows our customers to concentrate on delivering best-in-class products rather than trying to mitigate difficult design and certification issues.


Have another topic that you would like Qorvo experts to cover? Email your suggestions to the Qorvo Blog team and it could be featured in an upcoming post. Please include your contact information in the body of the email.

Explore the Qorvo Blog

About the Author

Qorvo Wi-Fi Experts

Four of our resident Wi-Fi product marketing, application, and R&D design experts, Mudar Aljoumayly, David Guo, Igor Lalicevic, and Jaidev Sharma, teamed up to create this blog post. Based on their collective knowledge of Wi-Fi device design, all have guided the advancement of filters for customers developing state-of-the-art Wi-Fi applications.

A Beginner’s Guide to Segmentation in Satellite Images: Walking through machine learning techniques for image segmentation and applying them to satellite imagery

A Beginner’s Guide to Segmentation in Satellite Images: Walking through machine learning techniques for image segmentation and applying them to satellite imagery

Blog from:

In my first blog, I walked through the process of acquiring and doing basic change analysis on satellite data. In this post, I’ll be discussing image segmentation techniques for satellite data and using a pre-trained neural network from the SpaceNet 6 challenge to test an implementation out myself.

What is image segmentation?

As opposed to image classification, in which an entire image is classified according to a label, image segmentation involves detecting and classifying individual objects within the image. Additionally, segmentation differs from object detection in that it works at the pixel level to determine the contours of objects within an image.


Image for post


In the case of satellite imagery, these objects may be buildings, roads, cars, or trees, for example. Applications of this type of aerial imagery labeling are widespread, from analyzing traffic to monitoring environmental changes taking place due to global warming.


Image for post


The SpaceNet project’s SpaceNet 6 challenge, which ran from March through May 2020, was centered on using machine learning techniques to extract building footprints from satellite images—a fairly straightforward problem statement for an image segmentation task. Given this, the challenge provides us with a good starting point from which we can begin to build understanding of what is an inherently advanced process.

I’ll be exploring approaches taken to the SpaceNet 6 challenge later in the post, but first, let’s explore a few of the fundamental building blocks of machine learning techniques for image segmentation to uncover how code can be used to detect objects in this way.

Convolutional Neural Networks (CNNs)

You’re likely familiar with CNNs and their association with computer vision tasks, particularly with image classification. Let’s take a look at how CNNs work for classification before getting into the more complex task of segmentation.

As you may know, CNNs work by sliding (i.e. convolving) rectangular “filters” over an image. Each filter has different weights and thus gets trained to recognize a particular feature of an image. The more filters a network has—or the deeper a network is—the more features it can extract from an image and thus the more complex patterns it can learn for the purpose of informing its final classification decision. However, given that each filter is represented by a set of weights to be learned, having lots of filters of the same size as the original input image makes training a model quite computationally expensive. It’s largely for this reason that filters typically decrease in size over the course of a network, while also increasing in number such that fine-grained features can be learned. Below is an example of what the architecture for an image classification task might look like:


Image for post


As we can see, the output of the network is a single prediction for a class label, but what would the output be for a segmentation task, in which an image may contain objects of multiple classes in different locations? Well, in such a case, we want our network to produce a pixel-wise map of classifications like the following:


Image for post

An image and its corresponding simplified segmentation map of pixel class labels. Source

To generate this, our network has a one-hot-encoded output channel for each of the possible class labels:


Image for post


These maps are then collapsed into one by taking the argmax at each pixel position.

The tricky part of achieving this segmentation is that the output has to be aligned with the input image—we can’t follow the exact same downsampling architecture that we use in a classification task to promote computational efficiency because the size and locality of the class areas must be preserved. The network also needs to be sufficiently deep to learn detailed enough representations of each of the classes such that it can distinguish between them. One of the most popular kinds of architecture for meeting these demands is what is known as a Fully Convolutional Network.

Fully Convolutional Networks (FCNs)

FCN’s get their name from the fact that they contain no fully-connected layers, that is, they are fully convolutional. This structure was first proposed by Long et al. in a 2014 paper, which I aim to summarize key points of here.

With standard CNNs, such as those used in image classification, the first layer of the network is fully-connected, meaning it has the same dimensions as the input image; this means that the size of the first layer must be fixed to align with the input image. Not only does this render the network inflexible to inputs of different sizes, it also means that the network uses global information (i.e. information from the entire image) to make its classification decision, which does not make sense in the context of image segmentation in which our goal is to assign different class labels to different regions of the image. Convolutional layers, on the other hand, are smaller than the input image so that they can slide over it—they operate on local input regions.

In short, FCNs replace the fully-connected layers of standard CNNs with convolutional layers with large receptive fields. The following figure illustrates this process. We see how a standard CNN for classification of a cat-only image can be transformed to output a heatmap for localizing the cat in the context of a larger image:


Image for post


Moving through the network, we can see that the size of the layers getting smaller and smaller for the sake of learning finer features in a computationally efficient manner—a process known as “downsampling.” Additionally, we notice that the cat heatmap is of coarser resolution than the input image. Given these factors, how does the coarse feature map get translated back to the size of the input image at a high enough resolution such that the pixel classifications are meaningful? Long et al. used what is known as learned upsampling to expand the feature map back to the same size as the input image and a process they refer to as “skip layer fusion” to increase its resolution. Let’s take a closer look at these techniques.

Demystifying Learnable Upsampling

Prior approaches to upsampling relied on hard-coded interpolation methods, but Long et al. proposed a technique that uses transpose convolution to upsample small feature maps in a learnable way. Recall the way that normal convolution works:


Image for post


The filter represented by the shaded area slides over the blue input feature map, computing dot products at each position to be recorded in the green output feature map. The weights of the filter are what is being learned by the network during training.

Transpose convolution works differently: the filter’s weights are all multiplied by the scalar value of the input pixel it is positioned over, and these values get projected to the output feature map. Where filter projections in the output map overlap, their values are added.


Image for post


Long et al. use this technique to upsample the feature map rendered by network’s downsampling layers in order to translate its coarse output back to pixels that align with those of the input image, such that the network’s architecture looks like this:


Image for post

An example of a final upsampling layer appended to the downsampling path to render a full-sized segmentation map. Note that the final feature map has 21 channels, representing the number of classes for the particular segmentation challenge being explored in the paper. Source

However, simply adding one of these transpose convolutional layers at the end of the downsampling layers yields spatially imprecise results, as the large stride required to make the output size match the input’s (32 pixels, in this case) limits the scale of detail the upsampling can achieve:


Image for post

The upsampled segmentation map (left) is appropriately scaled to the input image but lacks spatial precision. Source

Luckily, this lack of spatial precision can be somewhat mitigated by “fusing” information from layers with different strides, as we’ll now discuss.

Skip Layer Fusion

As previously mentioned, a network must be deep enough to learn detailed features such that it can make faithful classification predictions; however, zeroing in closely on any one part of an image comes at the cost of losing spatial context of the image as a whole, making it harder to localize your classification decision in the process of zooming back out. This is the inherent tension at play in image segmentation tasks, and one that Long et al. work to resolve using skip connections.

In neural networks, a skip connection is a fusion between non-adjacent layers; in this case, skip connections are used to transfer local information by summing feature maps from the downsampling path with feature maps from the upsampling path. Intuitively, this makes sense: with each step we take through the downsampling path of the network, global information gets lost as we zoom into a particular area of the image and the feature maps get coarser, but once we have gone sufficiently deep to make an accurate prediction, we wish to zoom back and localize it, which we can do utilizing information stored in the higher resolution feature maps from the downsampling path of the network. Let’s take a more in depth look at this process by referencing the architecture Long et al. use in their paper:


Image for post

Visualization of skip connections (left arrows) in a network and their effect on the granularity of resulting segmentation maps. Source (modified)

Across the top of the image is the network’s downsampling path, which we can see follows a pattern of two or three convolutions followed by a pooling layer. conv7 represents the coarse feature map generated at the end of the downsampling path, akin to the cat heatmap we saw earlier. The “32x upsampled prediction” is the result of the first architecture without any skip connections, accomplishing all of the necessary upsampling with a single transpose convolutional layer of a 32 pixel stride.

Let’s walk through the “FCN-16s” architecture, which involves one skip connection (see the second row of the diagram). Though it is not visualized, a 1×1 convolution layer is added on top of the “pool4” feature map to produce class predictions for all its pixels. But the network does not end there—it proceeds to downsample by a factor of 2 once more to produce the “conv7” class prediction map. Since the conv7 map is of half the dimensionality of the pool4 map, it is upsampled by a factor of 2 and its predictions are added to those of the pool4, producing a combined prediction map. This result is upsampled via a transpose convolution with a stride of 16 to yield the final “FCN-16s” segmentation map, which we can see achieves better spatial resolution than the FCN-32s map. Thus, although the conv7 predictions experience the same amount of upsampling in the end as in the FCN-32s architecture (given that 2x upsampling followed by 16x upsampling = 32x upsampling), factoring the predictions from the pool4 layer improves the result greatly. This is because pool4 reintroduces valuable spatial information from the input image into the equation—information that otherwise gets lost in the additional downsampling operation for producing conv7. Looking at the diagram, we can see that the “FCN-8s” architecture follows a similar process, but this time a skip connection is also added from the “pool3” layer, which we see yields an even higher fidelity segmentation map.

FCNs—Where to go from here?

FCNs were a big step in semantic segmentation for their ability to factor in both deep, semantic information and fine, appearance information to make accurate predictions via an “encoding and decoding” approach. But the original architecture proposed by Long et al. still falls short of ideal. For one, it results in somewhat poor resolution at segmentation boundaries due to loss of information in the downsampling process. Additionally, overlapping outputs of the transpose convolution operation discussed earlier can cause undesirable checkerboard-like patterns in the segmentation map, which we see an example of below:


Image for post

Criss-crossing patterns in a segmentation heatmap resulting from overlapping transpose convolution outputs. Source

Many models have built upon the promising baseline FCN architecture, seeking to iron out its shortcomings, “U-net” being a particularly notable iteration.

U-Net—An Optimized FCN

U-net was first proposed in a 2015 paper as an FCN model for use in biomedical image segmentation. As the paper’s abstract states, “The architecture consists of a contracting path to capture context and a symmetric expanding path that enables precise localization,” yielding a u-shaped architecture that looks like this:


Image for post

An implementation of the U-net architecture. Numbers at the top of the feature maps denote their number of channels, numbers at the bottom left denote their x-y-size. The white feature maps represent are copies from the downsampling path, which we can see get concatenated to feature maps in the upsampling path. Source

We can see that the network involves 4 skip connections—after each transpose convolution (or “up-conv”) in the upsampling path, the resulting feature map gets concatenated with one from the downsampling path. Additionally, we see that the feature maps in the upsampling path have a larger number of channels than in the baseline FCN architecture for the purpose of passing more context information to higher resolution layers.

U-net also achieves better resolution at segmentation boundaries by pre-computing a pixel-wise weight map for each training instance. The function used to compute the map places higher weights on pixels along segmentation boundaries. These weights are then factored into the training loss function such that boundary pixels are given higher priority for being classified correctly.

We can see that the original U-net architecture yields quite fine-grained results in its cellular segmentation tasks:


Image for post

U-net segmentations results in images and d, with ground truth boundaries outlined in yellow. Source

The development of U-net yet was another milestone in the field of computer vision, and five years later, models continue to expound upon its u-shaped architecture to achieve better and better results. U-net lends itself well to satellite imagery segmentation, which we will circle back to soon in the context of the SpaceNet 6 challenge.

Further Developments in Image Segmentation

We’ve now walked through an evolution of a few basic image segmentation concepts—of course, only scratching the surface of a topic at the center of a vast, rapidly evolving field of research. Here is a list of a few other interesting image segmentation concepts and applications, with links should you wish to explore them further:

  • Instance segmentation is a hybrid of object detection and image segmentation in which pixels are not only classified according to the class they belong to, but individual objects within these classes are also extracted, which is useful when it comes to counting objects, for example.
  • Techniques for image segmentation extend to video segmentation as well; for example, Google AI uses an “hourglass segmentation network architecture” inspired by U-net for real-time foreground-background separation in YouTube stories.
  • Clothing image segmentation has been used to help retailers match catalogue items with physical items in warehouses for more efficient inventory management.
  • Segmentation can be applied to 3D volumetric imagery as well, which is particular useful in medical applications; for example, research has been done on using it to monitor the development of brain lesions in stroke patients.
  • Many tools and packages have been developed to make image segmentation accessible to people of various skill levels. For instance, here is an example that uses Python’s PixelLib library to achieve 150-class segmentation with just 5 lines of code.

Now, let’s walk through actually implementing a segmentation network ourselves using satellite images and a pre-trained model from the SpaceNet 6 challenge.

The SpaceNet 6 Challenge

The task outlined by the SpaceNet challenge is to use computer vision to automatically extract building footprints from satellite images in the form of vector polygons (as opposed to pixel maps). In the challenge, predictions generated by a model are determined viable or not by calculating their intersection over union with ground truth footprints. The model’s f1 score over all the test images is calculated according to these determinations, serving as the metric for the competition.

The training dataset consists of a mix of mostly synthetic aperture radar (SAR) and a few electro-optical (EO) 0.5m resolution satellite images collected by Capella Space over Rotterdam, the Netherlands. The testing dataset contains only SAR images (for further explanation on SAR imagery, take a look at my last blog). The dataset being structured in this way makes the challenge particularly relevant to real-world applications, as SpaceNet explains, it is meant to “mimic real-world scenarios where historical optical data may be available, but concurrent optical collection with SAR is often not possible due to inconsistent orbits of the sensors, or cloud cover that will render the optical data unusable.”


Image for post

An example of a SAR image from the SpaceNet 6 dataset, with building footprint annotations shown in red. Source

More information on the dataset, including instructions for downloading it, can be found here. Additionally, SpaceNet released a baseline model, for which they provide explanation and code. Let’s explore the architecture of this model before implementing it to make predictions ourselves.

The Baseline Model


Image for post

TernausNet architecture. Source

The architecture SpaceNet uses as its baseline is called TernausNet, a variant of U-Net with a VGG11 encoder. VGG is a family of CNNs, VGG11 being one with 11 layers. TernausNet uses a slightly modified version of VGG11 as its encoder (i.e. downsampling path). The network’s upsampling path mirrors its downsampling path, with 5 skip connections linking the two. TernausNet improves upon U-Net’s performance by initializing the network with weights that were pre-trained on Kaggle’s Carvana dataset. Using a model pre-trained on other data can reduce training time and overfitting—an approach known as transfer learning. In fact, SpaceNet’s baseline takes advantage of transfer learning again by first training on only the optical portion of the training dataset, then using the weights it finds through this process as the initial weights in its final training pass on the SAR data.

Even with these applications of transfer learning, though, training the model on roughly 24,000 images is still a very time intensive process. Luckily, SpaceNet provides the weights for the model at its highest scoring epoch, which allow us to get the model up and running fairly easily.

Making Predictions from the Baseline Model

Step-by-step instructions for deploying the baseline model can be found in this blog. In short, the process involves spinning up an AWS Elastic Cloud Compute (EC2) instance to gain access to GPUs for more timely computation and loading the challenge’s Amazon Machine Image (AMI), which is pre-loaded with the software, baseline model and dataset. Keep in mind that the dataset is very large, so downloads may take some time.

Once your downloads are complete, you can find the PyTorch code defining the baseline model in takes care of image preprocessing and running training and testing operations. The weights of the pre-trained model with the best scoring epoch are found in the weights folder and are loaded when is run.

When we run an image through the model, it outputs a series of coordinates that define the boundaries of the building footprints we are looking to find as well as a mask on which these footprints are plotted. Let’s walk through the process of visualizing an image and its mask side-by-side to get a sense of how effective the baseline model is at extracting building footprints. Code for producing the following visualizations can be found here.

Getting a coherent visual representation of the SAR data is somewhat trickier than expected. This is because each pixel in a given image is assigned 4 values, corresponding to 4 polarizations of data in the X-band of the electromagnetic spectrum—HH, HV, VH and VV. In short, signals transmitted and received from a SAR sensor come in both horizontal and vertical polarization states, so each channel corresponds to a different combination of the transmitted and received signal types. These 4 channels don’t translate to the 3 RGB channels we expect for rendering a typical image. Here’s what it looks like when we select the channels one-by-one and visualize them in grayscale:


Image for post

Visual representations of the 4 polarizations of a single SAR image. Image by author

Notice that each of the 4 polarizations captures a slightly different representation of the same area of land. We can combine these representations to produce a single-channel span image to plot alongside the building footprint mask the model generated, which we convert to binary to make the boundaries more clear. With this, we can see that the baseline model did recognize the general shapes of several buildings:


Image for post

A visualization of the combined spectral bands of a SAR test image and the corresponding building footprint mask generated by the baseline model. Image by author

It is pretty cool to see the basic structures we’ve discussed in this post in action here, producing viable image segmentation results. But, it’s also clear that there is room for improvement upon this baseline architecture—indeed, it only achieves an f1 score of 0.21 on the test set.


The SpaceNet 6 challenge wrapped up in May, with the winning submission achieving an f1 score of 0.42—double that of the baseline model. More details on the outcomes of the challenge can be found here. Notably, all of the top 5 submissions implemented some variant of U-Net, an architecture that we now have a decent understanding of. SpaceNet will be releasing these highest performing models on GitHub in the near future and I look forward to trying them out on time series data to do some exploration with change detection in a future post.

Lastly, I’m very thankful for the thorough and timely assistance I received from Capella Space for writing this—their insight into the intricacies of SAR data as well as recommendations and code for processing it were integral to this post.



How To Benchmark ANN Algorithms– An Investigation Into The Performance Of Various Approximate Nearest-Neighbor Algorithms

Blog post from partner:

Approximate Nearest-Neighbors for new voter party affiliation. Credit:


The field of data science is rapidly changing as new and exciting software and hardware breakthroughs are made every single day. Given the rapidly changing landscape, it is important to take the appropriate time to understand and investigate some of the underlying technology that has shaped and will shape, the data science world. As an undergraduate data scientist, I often wish more time was spent understanding the tools at our disposal, and when they should appropriately be used. One prime example is the variety of options to choose from when picking an implementation of a Nearest-Neighbor algorithm; a type of algorithm prevalent in pattern recognition. Whilst there are a range of different types of Nearest-Neighbor algorithms I specifically want to focus on Approximate Nearest Neighbor (ANN) and the overwhelming variety of implementations available in python.

My first project with my internship at GSI Technology explored the idea of benchmarking ANN algorithms to help understand how the choice of implementation can change depending on the type and size of the dataset. This task proved challenging yet rewarding, as to thoroughly benchmark a range of ANN algorithms we would have to use a variety of datasets and a lot of computation. This would all prove to provide some valuable results (as you will see further down) in addition to a few insights and clues as to which implementations and implementation strategies might become industry standard in the future.

What Is ANN?

Before we continue its important to lay out the foundations of what ANN is and why is it used. New students to the data science field might already be familiar with ANN’s brother, kNN (k-Nearest Neighbors) as it is a standard entry point in many early machine learning classes.


Red points are grouped with the five (K) closest points.

kNN works by classifying unclassified points based on “k” number of nearby points where distance is evaluated based on a range of different formulas such as Euclidean distance, Manhattan distance (Taxicab distance), Angular distance, and many more. ANN essentially functions as a faster classifier with a slight trade-off in accuracy, utilizing techniques such as locality sensitive hashing to better balance speed and precision. This trade-off becomes especially important with datasets in higher dimensions where algorithms like kNN can slow to a grueling pace.

Within the field of ANN algorithms, there are five different types of implementations with various advantages and disadvantages. For people unfamiliar with the field here is a quick crash course on each type of implementation:


  • Brute Force; whilst not technically an ANN algorithm it provides the most intuitive solution and a baseline to evaluate all other models. It calculates the distance between all points in the datasets before sorting to find the nearest neighbor for each point. Incredibly inefficient.
  • Hashing Based, sometimes referred to as LSH (locality sensitive hashing), involves a preprocessing stage where the data is filtered into a range of hash-tables in preparation for the querying process. Upon querying the algorithm iterates back over the hash-tables retrieving all points similarly hashed and then evaluates proximity to return a list of nearest neighbors.
  • Graph-Based, which also includes tree-based implementations, starts from a group of “seeds” (randomly picked points from the dataset) and generates a series of graphs before traversing the graphs using best-first search. Through using a visited vertex parameter from each neighbor the implementation is able to narrow down the “true” nearest neighbor.
  • Partition Based, similar to hashing, the implementation partitions the dataset into more and more identifiable subsets until eventually landing on the nearest neighbor.
  • Hybrid, as the name suggests, is some form of a combination of the above implementations.

Because of the limitations of kNN such as dataset size and dimensionality, algorithms such as ANN become vital to solving classification problems with these kinds of constraints. Examples of these problems include feature extraction in computer vision, machine learning, and many more. Because of the prominence of ANN, and the range of applications for the technique, it is important to gauge how different implementations of ANN compare under different conditions. This process is called “Benchmarking”. Much like a traditional experiment we keep all variables constant besides the ANN algorithms, then compare outcomes to evaluate the performance of each implementation. Furthermore, we can take this experiment and repeat it for a variety of datasets to help understand how these algorithms perform depending on the type and size of the input datasets. The results can often be valuable in helping developers and researchers decide which implementations are ideal for their conditions, it also clues the creators of the algorithms into possible directions for improvement.

Open Source to the Rescue


Utilizing the power of online collaboration we are able to pool many great ideas into effective solutions

Beginning the benchmarking task can seem daunting at first given the scope and variability of the task. Luckily for us, we are able to utilize work already done in the field of benchmarking ANN algorithms. Aumüller, Bernhardsson, and Faithfull’s paper ANN-Benchmarks: A Benchmarking Tool for Approximate Nearest Neighbor Algorithms and corresponding GitHub repository provides an excellent starting point for the project.

Bernhardsson, who built the code with help from Aumüller and Faithfull, designed a python framework that downloads a selection of datasets with varying dimensionality (25 to nearly 28,000 dimensions) and size (few hundred megabytes to a few gigabytes). Then, using some of the most common ANN algorithms from libraries such as scikit-learn or the Non-Metric Space Library, they evaluated the relationship between queries-per-second and accuracy. Specifically, the accuracy was a measure of “recall”, which measures the ratio of the number of result points that are true nearest neighbors to the number of true nearest neighbors, or formulaically:


Intuitively recall is simply the correct predictions made by the algorithm, over the total number of correct predictions it could have made. So a recall of “1” means that the algorithm was correct in its predictions 100% of the time.

Using the project, which is available for replication and modification, I went about setting up the benchmark experiment. Given the range of different ANN implementations to test (24 to be exact), there are many packages that will need to be installed as well as a substantial amount of time required to build the docker environments. Assuming everything installs and builds as intended the environment should be ready for testing.



After three days of run time for the GloVe-25-angular dataset, we finally achieved presentable results. Three days of runtime was quite substantial for this primary dataset, however as we soon learned this process can be sped up considerably. The implementation of the benchmark experiment defaults to running benchmarks twice and averaging the results to better account for system interruptions or other anomalies that might impact the results. If this isn’t an issue, computation time could be halved by only performing the benchmark tests once each. In our case we wanted to match Bernhardsson’s results so we computed the benchmark with the default setting of two runs per algorithm which produced the following:

Our results (top) and Bernhardsson’s results (bottom):



My Results vs Bernhardsson’s Results

As you can see from the two side by side plots of algorithm accuracy vs algorithm query speed there are some differences between my results and Bernhardsson’s. In our case, there are 18 functions plotted as opposed to 15 in the other. This is likely because the project has since been updated to include more functions following Bernhardsson’s initial tests. Furthermore, the benchmarking was performed on a different machine to Bernhardsson’s which likely produced some additional variability.

What we do see which is quite impressive is that many of the same algorithms that performed well for Bernhardsson also performed well in our tests. This suggests that across multiple benchmarks there are some clearly well-performing ANN implementations. NTG-onng, hnsw(nmslib) and hnswlib all performed exceedingly well in both cases. Hnsw(nmslib) and hnswlib both belong to the Hierarchical Navigable Small World family, an example of a graph-based implementation for ANN. In fact, many of the algorithms tested, graph-based implementations seemed to perform the best. NTG-onng is also an example of a graph-based implementation for ANN search. This suggests that graph-based implementations of ANN algorithms for this type of dataset perform better than other competitors.

In contrast to the well-performing graph-based implementations, we can see BallTree(nmslib) and rpforest both of which in comparison are quite underwhelming. BallTree and rpforest are examples of tree-based ANN algorithms (a more rudimentary form of a graph-based algorithm). BallTree specifically is a hybrid tree-partition algorithm combining the two methods for the ANN process. It is likely a series of reasons that cause these two ANN algorithms to perform poorly when compared to HNSW or NTG-onng. However, the main reason seems to be that tree-based implementations execute slower under the conditions of this dataset.

Although graph-based implementations outperform other competitors it is worth noting that graph-based implementations suffer from a long preprocessing phase. This phase is required to construct the data structures necessary for the computation of the dataset. Hence using graph-based implementations might not be ideal under conditions where the preprocessing stage would have to be repeated.

One advantage our benchmark experiment had over Bernhardsson’s is our tests were run on a more powerful machine. Our machine (see appendix for full specifications) utilized the power of 2 Intel Xeon Gold 5115’s, an extra 32 GBs of DDR4 RAM totaling 64 GBs, and 960 GBs of solid-state disk storage which differs from Bernhardsson’s. This difference likely cut down on computation time considerably, allowing for faster benchmarking.

A higher resolution copy of my results can be found in the appendix.

Conclusion and Future Work


Further benchmarking for larger deep learning datasets would be a great next step.

Overall, my first experience with benchmarking ANN algorithms has been an insightful and appreciated learning opportunity. As we discussed above there are some clear advantages to using NTG-onng and hnsw(nmslib) on low dimensional smaller datasets such as the glove-25-angular dataset included with Erik Bernhardsson’s project. These findings, whilst coming at an immense computational cost, are none the less useful for data scientists aiming to tailor their use of ANN algorithms to the dataset they are utilizing.

Whilst the glove-25-angular dataset was a great place to start I would like to explore how these algorithms perform on even larger datasets such as the notorious deep1b (deep one billion) dataset which includes one billion 96 dimension points in its base set. Deep1b is an incredibly large file that would highlight some of the limitations as well as the advantages of various ANN implementations and how they trade-off between query speed and accuracy. Thanks to the hardware provided by GSI Technology this experiment will be the topic of our next blog.


  1. Computer specifications: 1U GPU Server 1 2 Intel CD8067303535601 Xeon® Gold 5115 2 3 Kingston KSM26RD8/16HAI 16GB 2666MHz DDR4 ECC Reg CL19 DIMM 2Rx8 Hynix A IDT 4 4 Intel SSDSC2KG960G801 S4610 960GB 2.5″ SSD.
  2. Full resolution view of my results:


  1. Aumüller, Martin, Erik Bernhardsson, and Alexander Faithfull. “ANN-benchmarks: A benchmarking tool for approximate nearest neighbor algorithms.” International Conference on Similarity Search and Applications. Springer, Cham, 2017.
  2. Liu, Ting, et al. “An investigation of practical approximate nearest neighbor algorithms.” Advances in neural information processing systems. 2005.
  3. Malkov, Yury, et al. “Approximate nearest neighbor algorithm based on navigable small-world graphs.” Information Systems 45 (2014): 61–68.
  4. Laarhoven, Thijs. “Graph-based time-space trade-offs for approximate near neighbors.” arXiv preprint arXiv:1712.03158 (2017).