<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Cyrus Behroozi on Medium]]></title>
        <description><![CDATA[Stories by Cyrus Behroozi on Medium]]></description>
        <link>https://medium.com/@cyrus.behroozi?source=rss-f3a2def729a9------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*LBiOmgwyQRtMh91h7VgTHA.jpeg</url>
            <title>Stories by Cyrus Behroozi on Medium</title>
            <link>https://medium.com/@cyrus.behroozi?source=rss-f3a2def729a9------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 15:20:42 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@cyrus.behroozi/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Pangiam Continues to Eliminate Bias in Face Recognition Algorithms]]></title>
            <link>https://medium.com/trueface-ai/pangiam-continues-to-eliminate-bias-in-face-recognition-algorithms-c85e58e07814?source=rss-f3a2def729a9------2</link>
            <guid isPermaLink="false">https://medium.com/p/c85e58e07814</guid>
            <category><![CDATA[face-recognition]]></category>
            <category><![CDATA[fairface]]></category>
            <category><![CDATA[bias]]></category>
            <category><![CDATA[trueface]]></category>
            <dc:creator><![CDATA[Cyrus Behroozi]]></dc:creator>
            <pubDate>Wed, 07 Sep 2022 17:05:24 GMT</pubDate>
            <atom:updated>2022-09-08T00:55:21.710Z</atom:updated>
            <content:encoded><![CDATA[<p>If you have followed our journey at Trueface, a Pangiam company, you’ll already know that eliminating bias in our face recognition algorithms is a top priority. We believe that machine learning algorithms should not discriminate against any ethnicity or minority group, and should have equitable performance across all demographic groups. To that end, <strong>we are pleased to report that our latest model, TFV5.1, achieved a large decrease in bias across all ethnicity and gender groups.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4Ok2NcE3dnKE_4EOuMkcww.jpeg" /></figure><p>This post is a continuation of two previous blog posts outlining our drive to quantify and eliminate bias. If you are new to Trueface, you can catch up <a href="https://medium.com/trueface-ai/trueface-tackles-facial-recognition-gender-and-ethnicity-bias-4c8b53de7806">here</a> and <a href="https://medium.com/trueface-ai/trueface-reduces-bias-across-all-ethnicities-in-latest-face-recognition-model-fe5509eda298">here</a>. In an effort to continue our transparency, I will share the results of the same evaluation on our newest model, TFV5.1. This is the very same model used in our most recent NIST FRVT 1:N submission, where we ranked as the 5th most accurate algorithm in the West in the Visa Kiosk category.</p><p>For this evaluation, we once again make use of the <a href="https://arxiv.org/abs/1908.04913">Fairface</a> dataset. The Fairface dataset contains a balanced number of face images from seven major ethnic groups and contains no more than a single image for each identity. In the evaluation, we generate a face recognition template for each image in the dataset, then compare every face template against one another to generate a similarity score. Generally, when quantifying the performance of a face recognition model, we generate and plot a Detection Error Tradeoff (DET) curve. However, since every comparison performed in our evaluation is an impostor match (a comparison of two different identities), we instead plot the False Positive Rate (FPR) vs similarity threshold. <strong>A lower and flatter curve indicates better performance because it means that there are fewer false positives at the given threshold</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*iM3z-IcKccRZJ096Jam4qw.png" /><figcaption>Comparison of bias in face recognition models</figcaption></figure><p>The plots above compare the bias in three of our face recognition models: TFV4, TFV5 , and TFV5.1 (you can expand the image by clicking on it). As can be seen on the charts, the green line (TFV5.1) is significantly lower than both the orange (TFV5) and blue (TFV4) lines on every subplot. This indicates that TFV5.1 has less bias when compared to TFV5 and TFV4 for all demographic and gender groups.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/1*zhBkB2f4_4CKBenCSN4otw.png" /><figcaption>Overall FPR against similarity threshold.</figcaption></figure><p>When we combine the above demographic and gender subplots to generate an overall plot of FPR, TFV5.1 has significantly better performance than the two preceding models for every similarity threshold value. Hence, the TFV5.1 algorithm provides us with better accuracy while also ensuring minimal bias across different groups.</p><p>As leaders in the computer vision industry, we have a responsibility to achieve parity of performance across all ethnicities and genders. The benefits of this powerful technology should be equitable. Let us state plainly that we will not stop until this goal has been achieved. We are excited to realize the advantages of this technology together.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c85e58e07814" width="1" height="1" alt=""><hr><p><a href="https://medium.com/trueface-ai/pangiam-continues-to-eliminate-bias-in-face-recognition-algorithms-c85e58e07814">Pangiam Continues to Eliminate Bias in Face Recognition Algorithms</a> was originally published in <a href="https://medium.com/trueface-ai">Trueface</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Trueface Achieves Most Accurate Face Recognition Software for Masked Faces at Biometric Rally]]></title>
            <link>https://medium.com/trueface-ai/trueface-achieves-most-accurate-face-recognition-software-for-masked-faces-at-biometric-rally-9647df80d381?source=rss-f3a2def729a9------2</link>
            <guid isPermaLink="false">https://medium.com/p/9647df80d381</guid>
            <category><![CDATA[computer-vision]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[trueface]]></category>
            <category><![CDATA[hd]]></category>
            <category><![CDATA[face-recognition]]></category>
            <dc:creator><![CDATA[Cyrus Behroozi]]></dc:creator>
            <pubDate>Tue, 17 May 2022 19:46:22 GMT</pubDate>
            <atom:updated>2022-05-17T19:46:22.807Z</atom:updated>
            <content:encoded><![CDATA[<h3>Trueface Achieves Most Accurate Face Recognition for Masked Faces at Biometric Rally</h3><p>U.S. Department of Homeland Security (DHS) Maryland Test Facility Biometric Rally 2021</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5ky5dIGpmGBIySYstcrWgg.jpeg" /></figure><p>The <a href="https://mdtf.org/Rally2021/Results2021?Length=0">results are finally in</a> for the 2021 Department of Homeland Security (DHS) Biometric Technology Rally. As promised in our <a href="https://medium.com/trueface-ai/maryland-test-facility-evaluates-trueface-at-annual-biometric-rally-5d6970542d50">previous blog post</a>, we trained a face recognition model which is optimized for masked face images. <a href="http://trueface.ai/"><strong>Trueface</strong></a><strong>, a </strong><a href="https://www.pangiam.com/"><strong>Pangiam</strong></a><strong> company, ranked as the #1 most accurate face recognition software for masked faces.</strong> In the mask face category, Trueface achieved an overall matching-focused True Identification Rate (TIR) of 98.4%. <strong>When face masks were removed, Trueface was able to achieve 100% accuracy across all genders, races, and skin tones</strong>, as summarized below:</p><ul><li>100% TIR for Male and Female demographic groups</li><li>100% TIR for Asian, Black, and White demographic groups</li><li>100% TIR for Darker Skin and Lighter Skin groups</li></ul><h4>Competition Overview</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7i4c1sFzrzgSopQXc4QE-w.jpeg" /><figcaption>Overview of the 2021 biometric technology rally</figcaption></figure><p>The 2021 Biometric Technology Rally evaluated state-of-the-art biometric identity verification systems for high throughput use cases. Unlike other facial recognition benchmarks like NIST FRVT, whose tests are run only on datasets of existing imagery, the MdTF Rallies collect data in controlled environments meant to reflect the real world. Examples include kiosks, e-gates, and other low friction image-capture systems. Results are therefore much more akin to real-world environments vs results generated in a lab/sandbox.</p><p>This year, the rally focused on testing how well biometric systems worked in two key areas:</p><ul><li>in the presence and absence of face masks</li><li>for people in different demographic groups</li></ul><p>It is also worth mentioning that the application process is highly competitive; only a handful of the best biometric providers in the industry are selected to participate.</p><h4>Results Breakdown: Masked Performance</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*nNTaYFBkL3Ta8KfqMNznTA.jpeg" /><figcaption>Key takeaways from the 2021 Biometric Technology Rally</figcaption></figure><p>One of the key findings of the report is that, on average, biometric matching systems experienced a 9% reduction in TIR when masks are introduced, identifying only 86% of all test subjects successfully. <strong>The Trueface algorithm, which has been trained to work even when the face is partially occluded, only experienced a 0.9% accuracy reduction</strong>.</p><h4>Results Breakdown: Operational Focus vs Matching Focus</h4><p>Additionally, the Rally sought to measure not only matching systems’ accuracy for self-reported gender and ethnicity of volunteers but also what, if any, were the measurable effects of acquisitions systems on overall performance. The Maryland Test Facility describes the difference in measurement results below:</p><blockquote>Operational Focused True Identification Rate (TIR) — the performance of an acquisition and matching system combination.</blockquote><blockquote>Matching Focused True Identification Rate (matching-TIR) — the performance of an acquisition and matching system combination, discounting any failures of the acquisition system to submit images for matching.</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/953/1*VM7Mm3WTfPGqBIMjEoVYlA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/952/1*PuUtjvCga7eQGW7hvR7pNQ.png" /><figcaption>Trueface True Identification Rate: Matching Focus vs Operational Focus Rally Results</figcaption></figure><p>As can be seen in the left graph above, Trueface achieves a perfect 100% TIR for every gender and ethnicity (with the Ouray acquisition system). Since 2017, we have made it our mission to minimize the implicit bias in facial recognition models (explore our journey to parity of performance with our self-reported Fair Face challenge results <a href="https://medium.com/trueface-ai/trueface-reduces-bias-across-all-ethnicities-in-latest-face-recognition-model-fe5509eda298">here</a> and <a href="https://medium.com/trueface-ai/trueface-tackles-facial-recognition-gender-and-ethnicity-bias-4c8b53de7806">here</a>). These results are an exciting acknowledgement that we are on the right track to achieving our goal of performance parity.</p><p>The graph on the right plots Trueface’s TIR for each ethnicity and gender inclusive of acquisition system failures. We applaud the Maryland Test Facility for testing both matching and acquisition systems together in real-world situations, shedding light on the fact that external factors clearly affect TIR.</p><p>We firmly believe that facial recognition will only be whole-heartedly adopted when the benefits of the technology can be realized by all, and these results prove we are on the right track to achieving equitable performance. We’d like to thank the Maryland Test Facility for including Trueface in these tests and look forward to next year’s rally.</p><p><em>Acknowledgment: “This publication is based upon work conducted under the U.S. Department of Homeland Security Cooperative Research and Development Agreement №20-TCBI-013.”<br>Disclaimer: The views and/or conclusions contained in this document are those of the author(s) and should not be interpreted as necessarily representing the official policies, either expressed or implied, of the U.S. Department of Homeland Security (DHS), and do not constitute a DHS endorsement of the equipment tested or evaluated.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9647df80d381" width="1" height="1" alt=""><hr><p><a href="https://medium.com/trueface-ai/trueface-achieves-most-accurate-face-recognition-software-for-masked-faces-at-biometric-rally-9647df80d381">Trueface Achieves Most Accurate Face Recognition Software for Masked Faces at Biometric Rally</a> was originally published in <a href="https://medium.com/trueface-ai">Trueface</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Introducing Version 1.0 of the Trueface SDK]]></title>
            <link>https://medium.com/trueface-ai/introducing-version-1-0-of-the-trueface-sdk-8cd231fe320e?source=rss-f3a2def729a9------2</link>
            <guid isPermaLink="false">https://medium.com/p/8cd231fe320e</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[sdk]]></category>
            <category><![CDATA[computer-vision]]></category>
            <category><![CDATA[trueface]]></category>
            <dc:creator><![CDATA[Cyrus Behroozi]]></dc:creator>
            <pubDate>Mon, 14 Mar 2022 05:52:23 GMT</pubDate>
            <atom:updated>2022-03-14T05:52:23.078Z</atom:updated>
            <content:encoded><![CDATA[<p>Featuring a whole lot of great updates</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/480/1*GkIa5ZZ9MXDFz0-g_bwI6Q.gif" /><figcaption>What’s that you say? Major SDK updates?</figcaption></figure><p>If you’ve been wondering why the Trueface team has been so quiet recently, it’s because the engineering team has been working diligently to release version 1.0 of our flagship SDK. This major update brings breaking changes to improve accuracy, speed, and overall SDK architecture. This blog post dives into said improvements. We’ll cover:</p><ul><li>A brand new face recognition model: TFV6</li><li>New TensorRT backend for GPU inference</li><li>A move to a stateless SDK</li><li>Improved passive spoof</li><li>Dynamic library support</li></ul><p>Let’s dive right in!</p><h3>TFV6: Our Most Accurate Model for Masked Faces</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/700/1*i6gzHlkExx92pEIMwlwTrg.gif" /><figcaption>Face recognition with masked faces</figcaption></figure><p>We have released a new TFV6 face recognition model which is our most accurate model for masked face images. This model has comparable CPU inference time and memory usage as our TFV5 model, but <strong>boasts a 24% improvement in accuracy for masked probe images</strong>, as reported by <a href="https://pages.nist.gov/frvt/html/frvt11.html">NIST FRVT</a>.</p><p>Elect to use this model in situations where it is anticipated that the probe image contains a masked face (for 1 to N search) or where one or both face images are masked (for 1 to 1 comparisons).</p><h3>Achieve Lower Latency &amp; Higher Throughput With TensorRT</h3><p>The GPU SDK has been modified to use TensorRT for inference. TensorRT is a high-performance GPU inference framework optimized for NVIDIA GPUs. <strong>This new backend option not only significantly reduces latency while increasing throughput, but also provides the user with more control over their design, such as allowing the user to choose their precision level, optimal batch size, and maximum GPU memory usage</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/605/1*eDpnRpCpv8JJ0zBQ1EqKxw.png" /><figcaption>Speed comparison of SDK V0 and V1</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/605/1*NN0-jy_8KGwkynjrB094NQ.png" /><figcaption>GPU memory comparisons of SDK V0 and V1</figcaption></figure><p>The above images show a comparison of the template generation time and GPU memory usage for our most accurate face recognition model with the previous and new versions of the SDK. As can be seen, <strong>the new TensorRT backend (denoted by SDK V1) has lower latency / higher throughput and uses less GPU memory for inference</strong>. When batching is used, we can achieve sub-millisecond average inference times. Consult the <a href="https://docs.trueface.ai/Benchmarks-0b648f5a0cb84badb6425a12697a15e5?utm_source=Trueface+Engineering&amp;utm_campaign=c1a2fa020f-_25_02_2021_COPY_02&amp;utm_medium=email&amp;utm_term=0_6f58e9bf3a-c1a2fa020f-&amp;mc_cid=c1a2fa020f&amp;mc_eid=UNIQID">benchmarks</a> page for a full breakdown of performance.</p><p>This backend is available for TFV6 and newer models. It is worth noting that it can take up to one minute or longer to generate the TensorRT engine file the first time inference is run (learn more <a href="https://reference.trueface.ai/cpp/dev/latest/usage/faq.html?utm_source=Trueface+Engineering&amp;utm_campaign=c1a2fa020f-_25_02_2021_COPY_02&amp;utm_medium=email&amp;utm_term=0_6f58e9bf3a-c1a2fa020f-&amp;mc_cid=c1a2fa020f&amp;mc_eid=UNIQID#what-is-the-tensorrt-engine-file-and-what-is-it-used-for">here</a>).</p><h3>SDK Architecture: A Move To Stateless</h3><p>This is probably the most significant change made to our API to date. The SDK has been refactored to remove state. In the previous architecture, the SDK would internally store the image and any subsequent inference operations would be applied to said image. This design had serious limitations and forced the user to be mindful of thread safety. In the new design, the SDK no longer stores images internally; instead, the preprocessed image must be passed to each inference function.</p><p>With respect to the SDK API, this means that Trueface::SDK::setImage() has been replaced by Trueface::SDK::preprocessImage() which returns a Trueface::TFImage containing the pre-processed CPU or GPU image. This Trueface::TFImage must then be passed to the various inference functions.</p><p>By removing state from the SDK, the CPU SDK is now completely threadsafe. Additionally, the user of the SDK can now employ more versatile design patterns. For example, the user can consume video from one thread, call the Trueface::SDK::preprocessImage method and enqueue the resulting Trueface::TFImage, and then from a separate thread, consume the pre-processed images from the queue for running inference. Consult the <a href="https://github.com/getchui/sample-apps/tree/master/cpp_sdk/cpp_dist_bundle/c%2B%2B?utm_source=Trueface+Engineering&amp;utm_campaign=c1a2fa020f-_25_02_2021_COPY_02&amp;utm_medium=email&amp;utm_term=0_6f58e9bf3a-c1a2fa020f-&amp;mc_cid=c1a2fa020f&amp;mc_eid=UNIQID">C++</a> and <a href="https://github.com/getchui/sample-apps/tree/master/python/python_dist_bundle?utm_source=Trueface+Engineering&amp;utm_campaign=c1a2fa020f-_25_02_2021_COPY_02&amp;utm_medium=email&amp;utm_term=0_6f58e9bf3a-c1a2fa020f-&amp;mc_cid=c1a2fa020f&amp;mc_eid=UNIQID">Python</a> sample code for a demonstration of how to use this new API.</p><h3>Improved Passive Spoof Model</h3><p>We have improved our passive spoof model in the SDK. As a refresher, passive spoof is able to detect presentation attacks, such as a malicious user holding a face image or face recording up to a camera in an attempt to gain access to a system. Unlike active spoof which requires multiple input images and specific action from the end user, <strong>passive spoof only requires a single unconstrained input image</strong>. Add passive spoof to your face recognition pipeline in order to protect against presentation attacks.</p><h3>Dynamic Library Support</h3><p>Some of you asked for this, and we listened! The SDK is now shipped as both a static and dynamic library for Linux. Linking against the dynamic library allows you to swap out the SDK with an upgraded version without needing to re-compile your application (so long as there were no breaking changes to the API).</p><p>Although upgrading from V0 to V1 will take some development effort on your end, we certainly hope that by now you can see why it is worthwhile. If you are an existing client and require support with this transition process, please feel free to open a support ticket <a href="https://support.trueface.ai/support/tickets/new">here</a> and our engineering team will be happy to assist.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8cd231fe320e" width="1" height="1" alt=""><hr><p><a href="https://medium.com/trueface-ai/introducing-version-1-0-of-the-trueface-sdk-8cd231fe320e">Introducing Version 1.0 of the Trueface SDK</a> was originally published in <a href="https://medium.com/trueface-ai">Trueface</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Trueface Achieves Number 1 Global Rank on NIST FRVT for Speed]]></title>
            <link>https://medium.com/trueface-ai/trueface-achieves-top-3-global-rank-on-nist-frvt-for-speed-ceca7215c33b?source=rss-f3a2def729a9------2</link>
            <guid isPermaLink="false">https://medium.com/p/ceca7215c33b</guid>
            <category><![CDATA[face-recognition]]></category>
            <category><![CDATA[computer-vision]]></category>
            <category><![CDATA[nist]]></category>
            <category><![CDATA[trueface]]></category>
            <category><![CDATA[biometrics]]></category>
            <dc:creator><![CDATA[Cyrus Behroozi]]></dc:creator>
            <pubDate>Thu, 30 Sep 2021 23:02:35 GMT</pubDate>
            <atom:updated>2022-08-26T20:43:32.342Z</atom:updated>
            <content:encoded><![CDATA[<h4>Search 1 billion identities in 1 second ⚡️</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/500/1*XraZB8ush79BPN86-JHkEQ.gif" /><figcaption>Searches so fast, you can feel it.</figcaption></figure><p>We are proud to report that Trueface has achieved the <strong>number one highest global ranking </strong>on the <a href="https://github.com/usnistgov/frvt/blob/nist-pages/reports/11/frvt_11_report_2022_06_27.pdf">NIST FRVT 11</a> competition in the template match speed category. When it comes to biometric identity verification, Trueface has the <strong>number one fastest software in the World</strong>. Trueface executes a biometric template comparison in as little as <strong>103 nanoseconds. </strong>To put that into perspective, that is 1,400,000 times faster than the blink of an eye.</p><h4>What exactly is template match time?</h4><p>Face recognition algorithms work by generating a biometric template from an input face image. The biometric template is a fixed-length list of numbers that mathematically describe certain attributes of the face. When two of these biometric templates are provided to the algorithm, it compares the templates to determine the similarity, and thus the probability that the identity of the two templates are the same. The time taken to perform this comparison is the match time.</p><h4>Why is a fast template match speed critical?</h4><p>The above example describes the process of 1 to 1 identity verification. In a 1 to 1 verification situation, we only compare two biometric templates against each other to obtain a similarity score.</p><p>However, more often than not, we are performing 1 to N identity verification. In these situations, we do not simply compare two biometric templates against each other, but instead, compare a biometric template against a database of potentially millions of biometric templates. In such cases, <strong>the template match comparison must be performed millions of times in order to determine if there was a single match in the database</strong>. Often, this search operation needs to be performed at least 30 times a second for a single 30 fps video camera. Once you add multiple camera feeds or extended video to the equation, it’s easy to see why match speed is so critical.</p><p>If we were to search every frame of a seven-minute video against a database of 1 million identities, Trueface would complete that search in 39 minutes. It would take our closest American competitor 1 hour and 7 minutes.</p><blockquote>In certain scenarios, 28 minutes of extra time can make all the difference.</blockquote><h4>How does the Trueface algorithm perform on different hardware?</h4><p>The Trueface matching algorithm has been optimized for speed on various platforms ranging from server-grade CPUs to embedded devices. The following table summarizes the 1 to N search benchmarks on various hardware platforms:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*QAbaRLFzrj5ka2lG6aKHcA.png" /><figcaption>Source: <a href="https://docs.trueface.ai/Benchmarks-0b648f5a0cb84badb6425a12697a15e5">https://docs.trueface.ai/Benchmarks-0b648f5a0cb84badb6425a12697a15e5</a></figcaption></figure><p>To highlight a few of the key points in the table above, the Trueface algorithm can search through 100 million identities on a single 10th generation Intel i-7 CPU in 873 ms. This can be scaled to achieve 1 billion identities per second if using a CPU cluster or alternatively a single CPU with many threads (ex. AMD Threadripper). On the embedded end, the Trueface algorithm can search through 1 million identities in 35 ms on an NVIDIA Jetson Xavier, and 192 ms on a Raspberry Pi 4. Hence, the Trueface algorithm can meet real-time requirements for large collections sizes even on embedded devices.</p><p><em>Disclaimer: Results shown from NIST do not constitute an endorsement of any particular system, product, service, or company by NIST. For more information, visit </em><a href="https://www.nist.gov/programs-projects/face-recognition-vendor-test-frvt-ongoing"><em>https://www.nist.gov/programs-projects/face-recognition-vendor-test-frvt-ongoing</em></a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ceca7215c33b" width="1" height="1" alt=""><hr><p><a href="https://medium.com/trueface-ai/trueface-achieves-top-3-global-rank-on-nist-frvt-for-speed-ceca7215c33b">Trueface Achieves Number 1 Global Rank on NIST FRVT for Speed</a> was originally published in <a href="https://medium.com/trueface-ai">Trueface</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How to Cross Compile OpenCV and MXNET for NVIDIA Jetson (AArch64 CUDA)]]></title>
            <link>https://medium.com/trueface-ai/how-to-cross-compile-opencv-and-mxnet-for-nvidia-jetson-aarch64-cuda-99d467958bce?source=rss-f3a2def729a9------2</link>
            <guid isPermaLink="false">https://medium.com/p/99d467958bce</guid>
            <category><![CDATA[opencv]]></category>
            <category><![CDATA[cuda]]></category>
            <category><![CDATA[mxnet]]></category>
            <category><![CDATA[nvidia-jetson]]></category>
            <category><![CDATA[cross-compile]]></category>
            <dc:creator><![CDATA[Cyrus Behroozi]]></dc:creator>
            <pubDate>Wed, 26 May 2021 05:37:20 GMT</pubDate>
            <atom:updated>2021-05-26T05:37:20.931Z</atom:updated>
            <content:encoded><![CDATA[<p>A technical step by step guide.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*BPY4lylEUfb7W9yWivW-8w.jpeg" /><figcaption>NVIDIA Jetson Xavier NX</figcaption></figure><p>Today I’ll be discussing how to cross compile popular computer vision and machine learning frameworks such as OpenCV and MXNET for AArch64 CUDA targets. The most popular example of this target is the NVIDIA Jetson, a GPU-enabled embedded System on Module (SOM). Despite the NVIDIA Jetson being widely used, I’ve found that there isn’t clear or sufficient documentation for cross compiling for this target, especially for novice programmers who may require a step-by-step guide. Therefore, I have put together this guide to help fill in some of those knowledge gaps.</p><p>Before we dive into all things technical, let us begin with a quick recap of cross compiling.</p><h4>What does it mean to cross compile?</h4><p>Cross compilation is the act of compiling code for a platform (often known as the target) other than the one on which the compiler is running (often know as the host).</p><h4>Why cross compile instead of native compile on the target?</h4><p>The most common targets for cross compilation are lightweight embedded devices. Although it is often possible to native compile on the target platform, doing so will be extremely slow when compared to cross compiling on a more powerful x86 device — the smaller memory, less powerful CPUs, and slower disks of embedded devices are to blame.</p><p>As an example, I went ahead and native compiled MXNET with CUDA support on the NVIDIA Jetson Xavier NX (their most powerful Jetson product) using the max power settings. The native compilation took a whopping 18 hours to complete! However, using an x86 laptop, I was able to cross compile the same library for the same target in less than an hour.</p><h4>What are some of the challenges with cross compiling?</h4><p>There are a few nuances associated with cross compiling which can make it more challenging than native compilation. For starters, the appropriate toolchain (compiler + linker + librarian + any other tools) must be installed for the cross compilation. Things become more complicated when the executable or library that is to be cross compiled has additional dependencies. These dependency libraries must either be cross compiled as well or installed for the specific target if available using a tool such as Multiarch (more on Multiarch <a href="https://wiki.debian.org/Multiarch/HOWTO">here</a>). Sometimes the target platform filesystem must even be copied onto the host machine.</p><h3>TL;DR</h3><p>If you want to skip the step-by-step guide and just view the final results, visit the following Github repository which contains the complete Dockerfiles, build scripts, and toolchain files discussed below.</p><p><a href="https://github.com/cyrusbehr/cuda-aarch64-cc-mxnet-opencv">cyrusbehr/cuda-aarch64-cc-mxnet-opencv</a></p><h3>Let’s Dive into the Step by Step Guide</h3><p>Many dependencies must be installed before the libraries can be cross compiled; therefore, it is best to work in a Docker container to avoid “dirtying” the local environment. The following steps give an overview of the Dockerfile which installs all the dependencies and build tools required for the build environment.</p><h4>Creating the build environment Dockerfile</h4><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/89ce5eefd969df1fbd817c74e76d92f7/href">https://medium.com/media/89ce5eefd969df1fbd817c74e76d92f7/href</a></iframe><p>The first step is to choose the base Docker image. We will be using the nvidia/cuda:10.2-cudnn7-devel-ubuntu18.04 image which comes installed with CUDA 10.2 and cudnn7. Note, we are using the devel image which extends the runtime image by adding the CUDA compiler toolchain, x86 CUDA header files, and x86 CUDA static libraries.</p><p>Sounds like we are ready to go? Unfortunately not. Although the above image contains the CUDA compiler that we require (nvcc), we still need the AArch64 CUDA toolkit which contains the AArch64 CUDA headers and AArch64 CUDA libraries (which we will install later).</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/4e0fe039b5b1d53736192fc8a13f19b4/href">https://medium.com/media/4e0fe039b5b1d53736192fc8a13f19b4/href</a></iframe><p>Next, we install a few of the build tools that will be used, including cmake, ccache, and aarch64-linux-gnu-g++ (<a href="https://packages.debian.org/sid/crossbuild-essential-arm64">installed</a> through thecrossbuild-essential-arm64 package). We also specify the toolchain file to be used for the cross compilation as an environment variable. The contents of the toolchain file are shown below.</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/61c066f3823305e1cb564e03b6d7a33e/href">https://medium.com/media/61c066f3823305e1cb564e03b6d7a33e/href</a></iframe><p>As can be seen, we specifyaarch-linux-gnu-gcc as the cross compiler and also instruct cmake to use this as the CUDA host compiler.</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/185ff16fb527513e217aa03dea05e318/href">https://medium.com/media/185ff16fb527513e217aa03dea05e318/href</a></iframe><p>As stated in their <a href="https://mxnet.apache.org/versions/1.4.1/install/build_from_source.html#math-library-selection">documentation</a>, MXNET relies on the BLAS (basic linear algebra subprograms) library for numerical computations. Although it supports several options, we will be using OpenBLAS which is best suited for embedded systems. We must therefore cross compile OpenBLAS by invoking the make command and specifying the cross compiler to be used.</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/7e57c6c04de939accd57f4b55242d5f5/href">https://medium.com/media/7e57c6c04de939accd57f4b55242d5f5/href</a></iframe><p>In this step, which is arguably the most important, we install the cuda-cross-aarch64 and cuda-cross-aarch64-10-2 packages which contain the AArch64 CUDA toolkit. You may notice that if you simply try running apt-get install -y cuda-cross-aarch64 on the base image, the package won’t be found. For this reason, we must manually add the debian packages using dpkg -i. Finally, we create a symbolic link so that libcublas.so is found at the expected location.</p><p>At this point, our Docker file is ready and can be built into a docker image. An instance of the image can then be run as a container, and we can proceed with cross compiling our libraries in this docker container.</p><h4>Cross Compiling OpenCV</h4><p>The following build script can be used to cross compile OpenCV in the Docker container.</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/a1146ec860d83eff303cea6f823c20d2/href">https://medium.com/media/a1146ec860d83eff303cea6f823c20d2/href</a></iframe><p>A few things to note:</p><ul><li>In order to build OpenCV with CUDA support, the OpenCV contrib package must be installed. In the cmake command, the path to the contrib package must be specified: -DOPENCV_EXTRA_MODULES_PATH=....</li><li>On line 14, I had to apply a hack to resolve the issue discussed <a href="https://github.com/opencv/opencv/issues/19986">here</a>.</li><li>The OpenCV provided toolchain file is used instead of the toolchain file that was added to the Docker image.</li><li>The -DCUDA_ARCH_BIN argument is used to specify the CUDA compute capability that will be supported (learn more about compute capability <a href="https://developer.nvidia.com/cuda-gpus">here</a>).</li></ul><h4>Cross Compiling MXNET</h4><p>The following build script can be used to cross compile MXNET:</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/51d95df0164fefa7d32073ec1de2b484/href">https://medium.com/media/51d95df0164fefa7d32073ec1de2b484/href</a></iframe><p>Here are a few things to note on the above:</p><ul><li>As per the bug described in this <a href="https://github.com/apache/incubator-mxnet/issues/20222">issue</a>, line 15 of mxnet/cpp-package/CMakeLists.txt must be removed.</li><li>Since OpWrapperGenerator.py is disabled as a result of the above fix, op.h must be pre-generated and copied to the destination directory before running the build.</li><li>The toolchain file which was copied into the Docker container is used.</li></ul><p>If you have followed along this far, you should have successfully compiled MXNET and OpenCV for the AArch64 CUDA target. At this point, you can write your main application and compile and link against the cross compiled libraries.</p><p>Here at Trueface, we automate the compilation of our Trueface SDK as part of our CI pipeline. We apply many of the techniques explained above in order to compile for our supported target platforms which include x86, x86 CUDA, AArch64, AArch32, and AArch64 CUDA. To learn more about what we are up to, <a href="http://www.trueface.ai/">check us out</a> or feel free to<a href="/service/mailto:hello@trueface.ai"> reach out</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=99d467958bce" width="1" height="1" alt=""><hr><p><a href="https://medium.com/trueface-ai/how-to-cross-compile-opencv-and-mxnet-for-nvidia-jetson-aarch64-cuda-99d467958bce">How to Cross Compile OpenCV and MXNET for NVIDIA Jetson (AArch64 CUDA)</a> was originally published in <a href="https://medium.com/trueface-ai">Trueface</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Deep Learning Framework Benchmarks: CPU Inference Speed and Memory Usage]]></title>
            <link>https://medium.com/trueface-ai/deep-learning-framework-benchmarks-cpu-inference-speed-and-memory-usage-9cc58f5dd220?source=rss-f3a2def729a9------2</link>
            <guid isPermaLink="false">https://medium.com/p/9cc58f5dd220</guid>
            <category><![CDATA[benchmark]]></category>
            <category><![CDATA[trueface]]></category>
            <category><![CDATA[inference-engine]]></category>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[openvino]]></category>
            <dc:creator><![CDATA[Cyrus Behroozi]]></dc:creator>
            <pubDate>Wed, 31 Mar 2021 22:14:08 GMT</pubDate>
            <atom:updated>2021-07-10T19:41:04.460Z</atom:updated>
            <content:encoded><![CDATA[<h3>How to Choose the Right Deep Learning Framework for Your Deployment: A Guide</h3><p>Machine learning algorithms have become pervasive in our modern times. Anywhere you turn you will find some form of machine learning being used to solve complex problems: from big data analytics to weather prediction to face recognition. With this recent rise of machine learning, the number of deep learning frameworks — tools for creating abstraction from the math and statistics involved in machine learning — available to developers has also increased. With so many options, choosing the right framework for your deployment needs can be difficult.</p><p>In this tutorial, we will begin by discussing the important metrics to consider when choosing a ML framework. Then we will dive into each framework and reveal the results of our own benchmark tests which should help you when deciding which framework is best for your specific use case.</p><p>There are two main classes of frameworks: those which are feature-rich and have been optimized for training, and those which are fast and lightweight and have been optimized for inference. Today, we will be reporting the measured benchmarks for some of those CPU inference-optimized frameworks. For the uninitiated, here is a brief summary of the difference between training and inference:</p><blockquote><strong>Training:</strong> Training refers to the process of teaching a model to learn from the data it sees.</blockquote><blockquote><strong>Inference:</strong> Inference refers to the process of using a trained machine learning algorithm to make a prediction.</blockquote><p>In a production deployment, the aim is to maximize the inference speed while minimizing resource usage. The various inference frameworks have different implementations in the backend, therefore, it is necessary to benchmark the frameworks to determine which is right for your deployment requirements.</p><h3>Benchmark Metrics</h3><p>With CPU inference frameworks, there are several metrics one must consider when selecting a framework:</p><h4>Latency</h4><p>Latency refers to the time taken to process one unit of data provided only one unit of data is processed at a time. In simpler terms, latency is the time one has to wait to get the inference result. Low latency is critical when designing real-time systems. For example, a system processing a video streaming at 30 FPS <em>must</em> have a latency of less than 33.3ms in order to be real-time. This is often a challenge because more accurate models are generally larger and have more parameters, which consequently increases the inference time. Since CPUs have low-latency cores and ultra-fast cache, we generally aim to minimize latency as opposed to throughput on CPUs (more on throughput below).</p><h4>Throughput</h4><p>Throughput refers to the number of data units processed in one unit of time. With respect to our video processing example, it would refer to the number of video frames we could process in a fixed amount of time for a batch size greater than 1. Batch size refers to the number of samples (or images in our example) which can be processed at the same time. We generally optimize for throughput on GPUs as they are massively parallel devices — a GPU core on its own is quite slow, but modern GPUs have thousands of cores capable of running in parallel. <a href="http://searchivarius.org/blog/difference-between-cpu-and-gpu-or-why-we-cannot-use-gpu-everything">Here</a> is a good resource if you want to learn why we optimize for throughput on GPUs instead of CPUs. Despite this, there are still ways to optimize for throughput on CPU that do not use batching, including the technique of running multiple inference instances in parallel while reducing the number of CPU threads each individual instance uses. More on this below.</p><h4>Memory Usage</h4><p>This one is self-explanatory. Memory usage refers to the amount of memory or RAM used by running inference. We must be mindful of memory usage for runtime environments with limited memory such as embedded devices. Additionally, we don’t want our framework to hog all the memory so that we ensure some memory is left for other tasks in our pipeline (ex. video decoding). This metric is particularly important if we plan on running multiple inference instances in parallel, as each instance will generally not re-use the same memory.</p><h4>CPU Usage</h4><p>This is related to the topic of latency and throughput. Inference frameworks generally do not run inference using only a single thread. They instead launch several threads to run different operations in parallel, consequently reducing the latency. In general, the more threads we use for inference (particularly with heavy models such as a ResNet100), the lower the latency. However, there is a limit to this. Eventually, the overhead of adding and managing new threads outweighs the provided increase in speed. Below, I’ve added a latency vs thread count chart for one of the machine learning pipelines we use at <a href="https://www.trueface.ai/">Trueface</a>:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/600/1*8HK0rLb9_6NWYJKBqgnssQ.png" /></figure><p>As can be seen, the graph follows an exponential decay pattern, with the greatest reduction in latency being experiences when moving from 1 thread to 2 threads. The significance of this is that we can actually enforce a reduced thread count in order to increase the CPU throughput. Consider an example where we have a CPU with 8 threads.</p><p><strong>Scenario 1: Latency Optimized:</strong><em> </em>In this scenario, we have 1 instance running inference using all 8 threads. Using the chart above, we can approximate the latency to be 75ms — given a single input image, inference can be performed in 75ms. Therefore, this scheme optimizes to reduce latency as much as possible. However, if the instance is provided with 100 input images to process, then it will take a total of 7.5s (100 images * 75ms ) to run inference.</p><p><strong>Scenario 2: Throughput Optimized: </strong>In this scenario, we have 8 instances running inference using only 1 thread each. Using the chart above, we can approximate the latency to be 400ms — given a single input image, inference will be performed in 400ms. Although this seems like a bad tradeoff compared to scenario 1, scenario 2 shines when we have many input samples. If the instances are provided with 100 input images to process, then it will take a total of 5s (100 images * 400ms / 8 instances) to run inference.</p><p>By running more instances in parallel and reducing the number of threads per instance, we have increased the latency but also increased the overall throughput in scenario 2. Inference frameworks allow the developer to dictate the number of threads used for inference, so it is therefore an important metric to track.</p><blockquote>Ultimately choosing to optimize for latency or throughput depends on the use case. For one-off inference such as using facial recognition for identity verification on a mobile device, you would want to optimize for latency so that the user gets feedback more quickly. For a use case such as scanning hundreds of hours of recorded video footage to find an individual of interest, optimizing for throughput is preferred.</blockquote><h4><strong>Ease of Installation and Deployment</strong></h4><p>This is an important metric if the software is to be shipped. At Trueface, we ship our inference code as part of a dependency-free compiled library that can run directly on the OS without requiring additional library installations. We must therefore be mindful of any dependencies which are incurred by adding the inference framework.</p><h3>Benchmark Test Setup</h3><p>The following benchmarks are running on unloadedDual Intel Xeon CPU E5-2630 v4 @2.20GHz (for those familiar with NIST FRVT, this is the same CPU used for their timing tests). The machine is equipped with 128Gb of RAM. The model used for inference is a ResNet100 model with 65.13 million params and 24.2 GFLOPs. The model input is a 112x112 RGB aligned face chip. The time it takes to read the image into memory and decode the image to an RGB image buffer is not included in the reported inference time. Any preprocessing required after this point to convert the decoded buffer to the expected framework-specific input is included in the inference time (including HWC to CHW conversions). The first inference time is discarded to ensure all network weights and params have been loaded. The number of inferences used to generate the average reported time is one the order of 1000.</p><h3>Frameworks</h3><p>In this benchmark test, we will compare the performance of four popular inference frameworks: MXNet, ncnn, ONNX Runtime, and OpenVINO. Before diving into the results, it is worth spending time to understand the contending frameworks.</p><h4>MXNet</h4><p>MXNet is generally <em>not</em> considered a high-performance deployment inference framework, but instead a training framework; the library is feature-rich for training but consequently has a lot of bloat for a deployment framework. MXNet was included in this benchmark as the model used in the benchmark was trained using this framework so it is a good baseline measurement. In the benchmark, we use the Intel Math Kernel Library for Deep Neural Networks (Intel MKL-DNN) backend, which takes advantage of Advanced Vector Extensions 2 (AVX2) for performing Single Instruction on Multiple pieces of Data (SIMD).</p><h4>ncnn</h4><p>This small and lesser-known library is powerful because, unlike the other frameworks, it is built as a static library and is dependency-free (other than requiring OpenMP). This makes it extremely easy to ship the library as part of your software. Although the library has been optimized for mobile devices and ARM CPUs, the library still boasts impressive results on X86_64 CPUs and keeps up with some of the larger frameworks. ncnn supports runtime CPU dispatching for AVX2 code paths.</p><h4><strong>ONNX Runtime</strong></h4><p>ONNX Runtime, developed by Microsoft, offers the most backend acceleration options of any framework, including MKL-DNN, OpenVINO, and TensorRT (CUDA GPU acceleration framework). It also supports different levels of model optimization. For this benchmark, we use the default backend with the maximum model optimization level enabled. With respect to deployment, the pre-compiled library can be conveniently downloaded and linked dynamically at runtime. To learn more about the benefits of ONNX, check out <a href="https://medium.com/trueface-ai/two-benefits-of-the-onnx-library-for-ml-models-4b3e417df52e">this</a> blog post written by one of my colleagues, a machine learning engineer at Trueface.</p><h4>OpenVINO</h4><p>Developed by the Intel team, this framework specifically targets Intel hardware. The framework is not only able to run inference on CPUs, but is also capable of running inference on specialized devices such as Intel GPUs, Intel VPUs (Neural Compute Stick), and even Intel FPGAs. Although the library does have several runtime dependencies, they are easy enough to download using the apt package manager on Ubuntu and bundled dependency installation scripts. Be mindful that in order to run inference with OpenVINO, the model must be converted to an Intermediate Representation (IR) format which may cause a slight difference in accuracy from the original model. For this benchmark, we use the MKL-DNN backend plugin.</p><h3>Results</h3><p>You can find the full inference code, library build scripts, and comprehensive results at the following github repository:</p><p><a href="https://github.com/getchui/deep_learning_inference_framework_benchmarks">getchui/deep_learning_inference_framework_benchmarks</a></p><p>The following graph plots the latency against the number of threads used:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/807/1*6iItodmG7R2QrSO5Xbfnow.png" /><figcaption>Latency vs number of threads</figcaption></figure><p>The following bar chart illustrates the latency of the contending frameworks when we focus on 8 threads:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/600/1*QrmmRrFVnFaDGJgj_Yusfg.png" /><figcaption>Latency at 8 threads</figcaption></figure><p>Both charts above show that OpenVINO had the lowest latency at all thread counts. Additionally, OpenVINO could be used to achieve the highest throughput. With respect to our examples above, OpenVINO performed the best in both scenario 1 and scenario 2 and would thus be the correct choice of framework.</p><p>The following chart summarizes the memory usage, reported as the maximum resident set size:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/600/1*sFkCw2VKP6y-TxN4M-dI8w.png" /><figcaption>Memory usage in Gb</figcaption></figure><p>ncnn used significantly more memory than the other frameworks. At the time of writing this blog, I opened an issue on the ncnn GitHub page to see if the library developers had any comments on this. You can follow the progress of that issue <a href="https://github.com/Tencent/ncnn/issues/2750">here</a>.</p><h3>Limitations</h3><ol><li>The benchmarks above are running on an Intel CPU. Unfortunately, we cannot assume inference framework performance or their respective rankings will be the same on AMD CPUs. Intel has <a href="https://software.intel.com/sites/products/documentation/doclib/daal/daal-user-and-reference-guides/daal_cpp_api/opt_notice.htm">admitted</a> to throttling some of their libraries including their Math Kernel Library (MKL) for non-intel CPUs. I ran a few tests on AMD CPUs — reported at the bottom of the <a href="https://github.com/getchui/deep_learning_inference_framework_benchmarks">linked Github repository</a> — but further investigation on non-intel CPUs must be performed.</li><li>ONNX Runtime and OpenVINO both modify the model. In doing so, the output and accuracy of the model are slightly different. Although the loss in accuracy appeared to be minimal during my testing, a more comprehensive test must be performed with a large dataset to quantify the exact loss in accuracy (perhaps by generating a Detection Error Tradeoff curve).</li></ol><h4>Summary</h4><p>We have defined the important metrics which are critical to understanding when deciding on an inference framework. As stated previously, in a production deployment, the aim is to maximize the inference speed while minimizing resource usage. We have also outlined the differences between latency and throughput and made note of why each is important when choosing a framework. Finally, we covered the popular inference frameworks before putting them through our own benchmark testing.</p><p>Congratulations! If you made it this far, you are well equipped to prioritize certain metrics and choose an ML framework that is best for your use case.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9cc58f5dd220" width="1" height="1" alt=""><hr><p><a href="https://medium.com/trueface-ai/deep-learning-framework-benchmarks-cpu-inference-speed-and-memory-usage-9cc58f5dd220">Deep Learning Framework Benchmarks: CPU Inference Speed and Memory Usage</a> was originally published in <a href="https://medium.com/trueface-ai">Trueface</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Trueface Reduces Bias Across all Ethnicities in Latest Face Recognition Model]]></title>
            <link>https://medium.com/trueface-ai/trueface-reduces-bias-across-all-ethnicities-in-latest-face-recognition-model-fe5509eda298?source=rss-f3a2def729a9------2</link>
            <guid isPermaLink="false">https://medium.com/p/fe5509eda298</guid>
            <category><![CDATA[bias]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[fairface]]></category>
            <category><![CDATA[face-recognition]]></category>
            <category><![CDATA[trueface]]></category>
            <dc:creator><![CDATA[Cyrus Behroozi]]></dc:creator>
            <pubDate>Fri, 26 Mar 2021 19:59:34 GMT</pubDate>
            <atom:updated>2021-03-26T19:59:34.371Z</atom:updated>
            <content:encoded><![CDATA[<h3>Trueface Reduces Bias Across all E<strong>thnicities</strong> with Newest Face Recognition Model</h3><p>At <a href="https://www.trueface.ai/">Trueface</a>, we are committed to eliminating bias in our face recognition models. Machine learning algorithms should not discriminate against any ethnicity or minority group. Our research and engineering teams are working diligently to ensure our models achieve parity of performance across all demographics.</p><p>If you require a refresher on algorithmic bias, check out this <a href="https://medium.com/trueface-ai/a-5-minute-tour-of-algorithmic-bias-mitigation-6bc731ede101">article</a> my colleague wrote which defines algorithmic bias and outlines a few of the strategies for mitigating bias in machine learning models.</p><p>In a previous <a href="https://medium.com/trueface-ai/trueface-tackles-facial-recognition-gender-and-ethnicity-bias-4c8b53de7806">article</a>, I introduced <a href="https://arxiv.org/abs/1908.04913">Fairface</a>, a unique dataset that contains roughly the same number of face images from each ethnicity and gender group. In that article, I quantified the bias in the Trueface production face recognition model. In an effort to continue our transparency, I will share the results of that very same evaluation on our <a href="https://reference.trueface.ai/cpp/dev/latest/usage/faq.html#what-are-the-differences-between-the-face-recognition-models">newest face recognition model</a>, TFV5.</p><p>The Fairface dataset contains a balanced number of face images from seven major ethnic groups and contains no more than a single image for each identity. In the evaluation, we generate a face recognition template for each image in the dataset, then compare every face template against one another to generate a similarity score. Generally, when quantifying the performance of a face recognition model, we generate and plot a Detection Error Tradeoff (DET) curve. However, since every comparison performed in our evaluation is an impostor match (a comparison of two different identities), we instead plot the False Positive Rate (FPR) vs similarity threshold. A lower and flatter curve indicates better performance because it means that there are fewer false positives at the given threshold.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-pB3dsX-kAWsTk2yHMLMng.png" /><figcaption>Comparison of bias in TFV4 vs TFV5</figcaption></figure><p>The plots above compare the bias in our TFV4 model to our recently released TFV5 model for each of the tested ethnicity and gender groups. <strong>TFV5 saw a decrease in bias across all ethnicity and gender groups.</strong> Of equal importance, TFV5 saw a significant reduction in bias in traditionally underrepresented groups such as East Asians and Southeast Asians. This is a result of supplementing our training dataset with additional ethically-sourced data from these underrepresented groups.</p><p>In general, we advise that our clients operate at a similarity score threshold of between 0.3 to 0.4, though the exact threshold is ultimately dictated by the desired False Positive Rate or False Negative Rate. What you will notice in the two plots below is that TFV5 has significantly fewer false positives in the operating region for all ethnicities. In a real world scenario of using single factor authentication for access control, TFV5’s performance improvement translates into increased security: halting more would-be intruders than our previous models. Of course, when paired with a second factor, like an RFID card or a PIN, security is increased even more.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/1*Th9choAmboUd8vrq-22abw.png" /><figcaption>TFV4 False Positive Rates by ethnicity</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/1*AIONAMB_70v_FUqxbHkcUw.png" /><figcaption>TFV5 False Positive Rates by ethnicity</figcaption></figure><p>At an operating threshold of 0.4, TVF5 correctly identifies 9,999 of every 10,000 identities. Raising the operating threshold will further reduce this False Positive Rate.</p><p>As leaders in the computer vision industry, we have a responsibility to achieve parity of performance across all ethnicities and genders. The benefits of this powerful technology should be equitable. Let us state plainly that we will not stop until this goal has been achieved. We are excited to realize the advantages of this technology together.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=fe5509eda298" width="1" height="1" alt=""><hr><p><a href="https://medium.com/trueface-ai/trueface-reduces-bias-across-all-ethnicities-in-latest-face-recognition-model-fe5509eda298">Trueface Reduces Bias Across all Ethnicities in Latest Face Recognition Model</a> was originally published in <a href="https://medium.com/trueface-ai">Trueface</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Trueface Achieves Top 3 Ranking in the West on NIST 1:N]]></title>
            <link>https://medium.com/trueface-ai/trueface-achieves-top-3-ranking-in-the-west-on-nist-1-n-db3c513eff60?source=rss-f3a2def729a9------2</link>
            <guid isPermaLink="false">https://medium.com/p/db3c513eff60</guid>
            <category><![CDATA[face-recognition]]></category>
            <category><![CDATA[computer-vision]]></category>
            <category><![CDATA[nist]]></category>
            <category><![CDATA[trueface]]></category>
            <category><![CDATA[identity-management]]></category>
            <dc:creator><![CDATA[Cyrus Behroozi]]></dc:creator>
            <pubDate>Wed, 24 Feb 2021 17:52:54 GMT</pubDate>
            <atom:updated>2022-04-06T15:02:13.038Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*9fAdA-onG1t0yH3QknPVEg.png" /><figcaption>NIST FRVT 1 to N false negative identification rates, ranked by visa kiosk category excluding China and Russia</figcaption></figure><p>We are proud to report our performance on <a href="http://www.trueface.ai/">Trueface’s</a> first submission to the National Institute of Standards and Technology (NIST) Face Recognition Vendor Test (FRVT) 1:N Identification. Out of 262 entries, we achieved the<strong> third-highest</strong> ranking by a western firm in the Visa Kiosk category.</p><p>As an extension of the <a href="https://medium.com/trueface-ai/trueface-scores-top-10-on-nist-facial-recognition-vendor-test-b5668d6d2000">NIST FRVT 1:1 competition</a>, the 1:N competition brings a greater challenge to developers. Whereas the 1:1 competition measures the effectiveness of an algorithm at comparing two provided face images to obtain a similarity score, the 1:N competition requires algorithms to enroll up to 12 million face recognition templates into a searchable database. Given a probe image, algorithms are evaluated on their speed and ability to search the database and find any potential identity matches.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*QE5_JztpsfxKxt9pkrIxIQ.png" /></figure><p>When one considers which NIST test to reference for their business the following example may be helpful: As referenced above, NIST 1:1 FRVT measures an algorithm’s ability to compare two provided face images, in order to verify the identity of an individual. 1:1 can be used to tell if a person presenting an ID or passport is who they claim to be. On the other hand, NIST 1:N is measuring an algorithm’s ability to find potential face matches from a reference database, thereby identifying one person out of many. Identification can be used to authenticate a person for use cases like granting access to a building or secure area.</p><p>Below is a summary of the performance of all algorithms listed in the main NIST FRVT 1:N Identification <a href="https://pages.nist.gov/frvt/html/frvt1N.html">leaderboard</a>. The box plots show the distribution of False Negative Identification Rates (FNIR) at a threshold used to achieve a False Positive Identification Rate (FPIR) of 0.003, with a lower FNIR indicating better performance.</p><blockquote>FNIR is the proportion of mated searches failing to return the mate above threshold. FPIR is the proportion of non-mated searches producing one or more candidates above threshold…The use of thresholding supports use of face recognition in making mostly automated decisions e.g. for access into facility.</blockquote><p>The Trueface algorithm performance is shown by the magenta dotted lines, and the rank below the plots refers to the Trueface algorithm rank in that category.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/898/1*nOy_LcMoClwkxrrMIJ7Dww.png" /><figcaption>Distribution of FNIR at FPIR of 0.003, Identification Mode.</figcaption></figure><p>We performed particularly well on the Mugshot Profile and Visa Kiosk categories, where we placed 16th out of 189 submissions and 15th out of 133 submissions, respectively. When submissions are scrubbed to reflect those based in <strong>Western countries only</strong>, Trueface ranks <strong>third</strong> in both these categories.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/978/1*StJuv1DZ2AKXxKAZdSwliQ.png" /><figcaption>Sample kiosk image (left) and mugshot profile image (right)</figcaption></figure><p>As you can see in the reference images above, these two categories are particularly challenging because they focus on face images where the subject is not looking directly at the camera, resulting in each face having extreme yaw and pitch angles. These image categories directly reflect real-world environments where perfect front-facing image data are the rare exception to the norm.</p><p>Whereas many of the other algorithms chose to ignore the Mugshot profile category entirely and thus performed very poorly in that category, Trueface has trained its face recognition model with face images at various pitch and yaw angles to ensure we have a very robust model. Thus, we are confident in our algorithm’s ability to excel in real-world applications where poor angles, sub-adequate lighting conditions, and other factors can diminish the performance of other models.</p><p>We’re proud of achieving a top-three ranking amongst western competitors with our very first submission to NIST FRVT 1:N. However, we’re not here to rest on our laurels. We are throwing this proverbial wreath in the backseat and continuing to blaze the trail to achieving fast, accurate, secure, and above all, unbiased facial recognition for all. We expect to improve with our subsequent submissions and continue to drive this industry forward.</p><p><em>Disclaimer: Results shown from NIST do not constitute an endorsement of any particular system, product, service, or company by NIST. For more information, visit </em><a href="https://www.nist.gov/programs-projects/face-recognition-vendor-test-frvt-ongoing"><em>https://www.nist.gov/programs-projects/face-recognition-vendor-test-frvt-ongoing</em></a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=db3c513eff60" width="1" height="1" alt=""><hr><p><a href="https://medium.com/trueface-ai/trueface-achieves-top-3-ranking-in-the-west-on-nist-1-n-db3c513eff60">Trueface Achieves Top 3 Ranking in the West on NIST 1:N</a> was originally published in <a href="https://medium.com/trueface-ai">Trueface</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Maryland Test Facility Evaluates Trueface at Annual Biometric Rally]]></title>
            <link>https://medium.com/trueface-ai/maryland-test-facility-evaluates-trueface-at-annual-biometric-rally-5d6970542d50?source=rss-f3a2def729a9------2</link>
            <guid isPermaLink="false">https://medium.com/p/5d6970542d50</guid>
            <category><![CDATA[biometrics]]></category>
            <category><![CDATA[trueface]]></category>
            <category><![CDATA[government]]></category>
            <category><![CDATA[hd]]></category>
            <category><![CDATA[face-recognition]]></category>
            <dc:creator><![CDATA[Cyrus Behroozi]]></dc:creator>
            <pubDate>Fri, 05 Feb 2021 22:11:18 GMT</pubDate>
            <atom:updated>2021-02-05T22:12:56.959Z</atom:updated>
            <content:encoded><![CDATA[<h3>Maryland Test Facility evaluates Trueface at 2020 Biometrics Rally measuring unmasked and masked identity performance</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Ho2BWiGJSNVCb7QAIEKDZQ.png" /><figcaption>Trueface was chosen to participate at the 2020 MdTF 2020 Rally.</figcaption></figure><p>Each year, the U.S Department of Homeland Security (DHS) Science and Technology Directorate (S&amp;T) hosts Biometric Technology Rallies. These rallies engage the biometrics industry and are meant to accelerate understanding of biometric and identity technologies, thus driving efficiencies and adoption by the DHS. Unlike other facial recognition benchmarks like NIST FRVT, LFW, and Megaface Challenge whose tests are run only on datasets of existing imagery, the MdTF Rallies collect data in the real world. Results are therefore much more akin to real-world environments vs results generated in a lab/sandbox.</p><p>In 2018, the rally tested biometric image (acquisition) systems, like podiums, booths and e-gates on their ability to capture images of participants. The 2019 rally tested acquisition and biometric matching algorithms (matching systems) on their combined ability to capture images and identify individuals.</p><p>In 2020, both acquisition and matching systems were tested but due to COVID-19 and the resulting changes to how we travel, matching systems were also judged on accuracy when subjects were wearing masks. <a href="http://www.trueface.ai/">Trueface</a> was selected to participate in the 2020 Biometric Technology Rally as a face recognition matching system.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Ng60UA2hxRX0OkGpG82wKg.png" /><figcaption>Overview of 2020 Biometrics Rally. Numbers shown in the graphic summarize median performance across all participants.</figcaption></figure><p>The 2020 rally was designed to emulate a high-throughput scenario such as a customs checkpoint where travelers debarking from an international flight or cruise could have their identity verified quickly while filling their customs application at an automated kiosk.</p><h4>Trueface Results: Non-Masked Faces</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/590/1*etWeZgd69SxIlDnCn4Ix2Q.png" /><figcaption>Trueface TIR using different acquisition systems on unmasked faces.</figcaption></figure><p>From a total of 544 <strong>non-masked</strong> face images provided from the codename “Stone” acquisition system, Trueface correctly identified 534, giving a TIR (True Identification Rate) of 98.2%.</p><h4>Trueface Results: Masked Faces</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/616/1*PKe_ZIJBbjh_nsmdDY84xg.png" /><figcaption>Trueface TIR using different acquisition systems on masked faces.</figcaption></figure><p>From a total of 408 <strong>masked</strong> face images provided from the codename “Round” acquisition system, Trueface correctly identified 386, giving a TIR (True Identification Rate) of 94.6%.</p><p>The Trueface model did not experience much of a reduction in accuracy for masked face images, especially in comparison to some of the competitors’ models. Trueface significantly outperformed the median masked face performance of 77% on several of the acquisition systems, demonstrating that the Trueface algorithm is quite effective, even with obstructions to the face.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/480/1*d_sIv1wRNgVCdygVyeREbw.gif" /><figcaption>Identification even with high obfuscation.</figcaption></figure><p>The Maryland Test Facility used a Trueface production model that had yet to be trained on masked-faces so we expect the performances we have discussed here to improve in the future. Trueface is continuously working towards improving its face recognition models in our ever-evolving social environment. The team is placing great emphasis on training with underrepresented demographics to <a href="https://medium.com/trueface-ai/a-5-minute-tour-of-algorithmic-bias-mitigation-6bc731ede101?source=friends_link&amp;sk=ed7b5869be88f14bd4c3d9844a4f5378">further reduce bias</a>.</p><p><em>Acknowledgment: “This publication is based upon work conducted under the U.S. Department of Homeland Security Cooperative Research and Development Agreement №20-TCBI-013.”<br>Disclaimer: The views and/or conclusions contained in this document are those of the author(s) and should not be interpreted as necessarily representing the official policies, either expressed or implied, of the U.S. Department of Homeland Security (DHS), and do not constitute a DHS endorsement of the equipment tested or evaluated.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5d6970542d50" width="1" height="1" alt=""><hr><p><a href="https://medium.com/trueface-ai/maryland-test-facility-evaluates-trueface-at-annual-biometric-rally-5d6970542d50">Maryland Test Facility Evaluates Trueface at Annual Biometric Rally</a> was originally published in <a href="https://medium.com/trueface-ai">Trueface</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Plug & Play Camera Provides Frictionless Temperature detection and Access Control]]></title>
            <link>https://medium.com/trueface-ai/plug-play-camera-provides-frictionless-temperature-detection-and-access-control-aea534b20bba?source=rss-f3a2def729a9------2</link>
            <guid isPermaLink="false">https://medium.com/p/aea534b20bba</guid>
            <category><![CDATA[covid-screening]]></category>
            <category><![CDATA[trueface]]></category>
            <category><![CDATA[elevated-body-temperature]]></category>
            <category><![CDATA[iryx]]></category>
            <category><![CDATA[computer-vision]]></category>
            <dc:creator><![CDATA[Cyrus Behroozi]]></dc:creator>
            <pubDate>Tue, 12 Jan 2021 16:59:24 GMT</pubDate>
            <atom:updated>2021-01-12T16:59:24.476Z</atom:updated>
            <content:encoded><![CDATA[<h3>Plug &amp; Play Camera for a Covid-World: Frictionless elevated body temperature detection and access control</h3><p>Trueface partners with iryx to deliver contactless temperature checks, frictionless identity verification, and spoof detection in a single camera.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/635/1*u1XHOFgxuaBa0T1ig-qCqQ.gif" /><figcaption>Temperature Screening and Mask Detection Demo</figcaption></figure><p>Living through a global pandemic has certainly been an interesting experience; it has demonstrated the essence of human resilience. We have been quick to adapt, allowing us to return to a pseudo-normal life. One of the things I now commonly see is businesses screening their patrons for Covid using a handheld temperature gun. Although this is a step in the right direction, there are a few shortcomings with this approach. The first is that this process is highly inefficient. Although this manual form of screening works well enough for small businesses like barbershops, it quickly breaks down in situations where you need to process a large volume of people such as in an airport. Second, this approach exposes the person administering the screening to potentially sick individuals. These risks are needless and can be avoided with the right technology.</p><p><a href="http://www.trueface.ai/">Trueface</a> has partnered with <a href="http://iryx.io/">iryx</a> to develop automated elevated body temperature screening technology that ensures we make it through this pandemic safely. Iryx is an American-based company that specializes in AI-enabled thermal cameras. Their flagship thermal camera, <a href="https://iryx.io/sensors/">the Indoor Mini</a>, is smaller than a baseball, yet features an advanced onboard computer vision engine that can perform machine learning inference at the edge.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/350/1*ngXhtDH1F2ZWKHrmK2JlgA.png" /><figcaption>iryx Indoor Mini</figcaption></figure><p>Leveraging the onboard machine learning engine, Trueface has been able to run many of its top-tier computer vision models directly onboard the camera in real-time. The Trueface <a href="https://docs.trueface.ai/ebt-docs">Elevated Body Temperature solution</a> uses machine learning to detect the location of the inner canthus of the eye then extracts the maximum temperature in this region. Several <a href="https://www.researchgate.net/publication/303738021_MEASURING_HUMAN_BODY_TEMPERATURE_USING_INFRARED_IN_THE_FOREHEAD_AND_INNER_CANTHI_OF_THE_EYE_VERSUS_TYMPANIC_FOREHEAD_AND_AXILLA_THERMOMETERS">studies</a> have shown that the inner canthus is the most accurate way to frictionlessly measure body temperature from a thermal image and its output correlates well to internal body temperature.</p><blockquote>The most reliable, repeatable and fastest method was the inner canthus of the eye in frontal facial thermal images. It presented good correlation with current ‘clinical gold standard tympanic thermometer and debatable forehead methods.</blockquote><p>The Trueface Elevated Body Temperature (EBT) solution is truly plug and play, and requires zero coding knowledge to set up and get started. Simply plug the Power over Ethernet (PoE) cable into the camera and the solution is ready.</p><p>For those wanting to delve a bit deeper and take advantage of more features, Trueface offers a rich <a href="https://docs.trueface.ai/ebt-docs">developer API</a> for the EBT solution. This includes a <a href="https://documenter.getpostman.com/view/12009415/T17M7RNe?version=latest">configuration server</a> as well as a <a href="https://docs.trueface.ai/ebt-websocket-iryx">websocket</a> that streams information such as the temperature, face mask label, and liveness information. Moreover, the API can be used to run Trueface’s face recognition models directly on the camera. This is an ideal solution for enterprises who require a frictionless method to not only verify the identities of those entering its premises but also certify they do not have an elevated temperature. This solution also scales easily, as the processing is performed directly on the camera itself. There’s no need to pay for beefy servers as more cameras are added.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/635/1*-_BTJHa8mi2WH9AuayPdAA.gif" /><figcaption>Face Recognition and Temperature Screening Demo</figcaption></figure><p>One of the most useful features accessible from the developer API is the built-in spoof detection which comes with the Trueface EBT solution. All face recognition systems are susceptible to presentation attacks, meaning a malicious actor could present a physical picture or a phone with a face image to the camera, impersonating a different identity in order to gain access to a digital or physical environment. While rare, presentation attacks are more common in remote access locations where implementing traditional security measures have not been realistic in the past. The Trueface EBT solution easily detects such attacks.</p><p>The passive thermal spoof detection approach we take is superior to standard conventional visual-based approaches. While passive visual-based approaches may look for the <a href="https://en.wikipedia.org/wiki/Screen-door_effect">screen door effect</a> or other spoof artifacts, the thermal approach compares the temperature profile of the detected face to the expected temperature profile of a human face. Simply put, it is much more difficult to spoof this sort of system.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/635/1*BKoq6rNBuTgef3aZFTflWw.gif" /><figcaption>Spoof Detection Demo</figcaption></figure><p>We recently pushed out V2.XX of the EBT solution, which comes paired with a firmware upgrade for the iryx camera. This new version adds a Trueface web interface which can be used for seamlessly installing and managing the Trueface EBT application.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/879/1*EE4K4YvNtn8cS0GkRfAsjw.png" /><figcaption>V2.XX Web Interface</figcaption></figure><p>In our ever-evolving COVID world, contactless capabilities have become imperative for enterprises and government agencies alike. From frictionless temperature measurement to identity verification to spoof detection, the iryx+Trueface solution is well-equipped to answer the dynamic requirements of today’s uncertain landscape.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=aea534b20bba" width="1" height="1" alt=""><hr><p><a href="https://medium.com/trueface-ai/plug-play-camera-provides-frictionless-temperature-detection-and-access-control-aea534b20bba">Plug &amp; Play Camera Provides Frictionless Temperature detection and Access Control</a> was originally published in <a href="https://medium.com/trueface-ai">Trueface</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>