<?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 Huma Lodhi on Medium]]></title>
        <description><![CDATA[Stories by Huma Lodhi on Medium]]></description>
        <link>https://medium.com/@huma.lodhi?source=rss-cf7c365df63f------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*dmbNkD5D-u45r44go_cf0g.png</url>
            <title>Stories by Huma Lodhi on Medium</title>
            <link>https://medium.com/@huma.lodhi?source=rss-cf7c365df63f------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sat, 21 Nov 2026 16:14:35 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@huma.lodhi/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[Football Event Detection: A Deep Learning Success Story]]></title>
            <link>https://medium.com/@huma.lodhi/football-event-detection-using-deep-learning-d1575b36f7b8?source=rss-cf7c365df63f------2</link>
            <guid isPermaLink="false">https://medium.com/p/d1575b36f7b8</guid>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[sports]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[football]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Huma Lodhi]]></dc:creator>
            <pubDate>Fri, 13 Aug 2021 14:49:31 GMT</pubDate>
            <atom:updated>2021-08-13T16:32:44.681Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>Paper Review: Automatic Pass Annotation from Soccer Video Streams Based on Object Detection and LSTM (Danilo Sorano, Fabio Carrara, Paolo Cintia, Fabrizio Falchi, and Luca Pappalardo)􀀀</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*UZz4I4cSk3zyR2RQ" /><figcaption>Photo by <a href="https://unsplash.com/@jesusance?utm_source=medium&amp;utm_medium=referral">Thomas Serer</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*k-t7DyzXqJyw2o-l" /><figcaption>Photo by <a href="https://unsplash.com/@universaleye?utm_source=medium&amp;utm_medium=referral">Munro Studio</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*y-Ip-QXNJbxgguPu" /><figcaption>Photo by <a href="https://unsplash.com/@tetrakiss?utm_source=medium&amp;utm_medium=referral">Arseny Togulev</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>Sports keep us healthy and, also stimulate our brain activities. These have also played an important role in the development of Artificial Intelligence and Machine Learning. Chess has been an attractive application domain since the early days of AI due to intelligence and reasoning required to play this.</p><p>Recently there has been an interest to apply AI and more specifically deep learning techniques to football for generating solutions for tasks ranging from game planning to detecting interesting event. Historically events like goals, key passes, saved shots are identified by human editors where this process is not time and resource effective. Sorano et. al., has presented PassNet, based on deep learning models, to automate pass detection for video streams.</p><p>In Football a pass is viewed as an event given by kicking of the ball by a player to another player where both players belong to the same team. An automated pass detection is considered as identification of a sequence of frames comprising the event by a computational system.</p><p>The automatic detection of passes is performed by decomposing this into three processes:</p><ul><li>extraction of features from video frames</li><li>detection of players and ball</li><li>categorisation of a video frame as a pass or otherwise</li></ul><p><strong>Feature Extraction</strong></p><p>In this module videos are decomposed into frames and then informative features are extracted for each frame. Feature are generated by using ResNet18 which is a deep Convolutional Neural network with 18 layers. ResNet is developed within residual learning framework and the structure of the network involves a shortcut connection to a pair of 3x3 filters. Once features are generated these are combined into a sequence of vectors. In this way high dimensional video frames are mapped into informative low dimensional feature vectors. For example, a frame of size 3 X 352 X 240 is mapped into a vector with 536 values.</p><p><strong>Players and Ball Detector</strong></p><p>You Only Look Once (YOLO) is an instance of Convolutional Neural Networks which give state-of-the-art performance for object detection. In this framework objects are detected in an image by formulating this as a regression problem, where features are extracted from the whole image and predictions are generated for each object / bounding box in the image. An image is viewed as an even grid where each cell of the grid predicts a given number of bounding boxes with confidence values and class probabilities. Final predictions are generated by multiplying probability values with confidence numbers. There are six well-known versions of YOLO, namely, YOLOv1, YOLOv2, YOLOv3, YOLOv4, PP-YOLO, YOLOv5.</p><p>In this work pre-trained YOLOv3 is used to detect players and the ball. Pre-training is conducted using COCO dataset which various image classes including images for balls and persons. YOLOv3 is provided with video frames, and it identifies presence of the desired objects in the frame. Once output is generated a frame is mapped into a 24-dimensional vector which represent the ball and the five closest players to it. Each object is represented as a four-dimensional vector. The first and the second elements of each individual vector are Boolean and tells the presence of the ball or a player and the occurrence of the object in the frame respectively. The next two values give the position of the ball or a player. The six vectors are combined to get a frame vector with 24 values and is termed Object position vector (OPV).</p><p><strong>Frames Classification</strong></p><p>An informative representation obtained by merging OPV with the vector given by ResNe18. The final module performs frames categorization using a deep network comprising a Bidirectional LSTM and two dense layers. Input to network is a 536-dimensional vector and the output is a Boolean value which tells the occurrence of a pass in a frame.</p><h3>Discussion</h3><p>The result achieved by PassNet for four football matches in the Italian first division are encouraging. The experiments are performed using two scenarios listed below:</p><p>a) Train PassNet on first half of the match and test on second half of the same match</p><p>b) Train PassNet on first halves of a some matches and test on second halves of the matches which are not used for training</p><p>The performance of PassNet is better for scenario a) as compared to scenario b). This shows limitation of the approach which may be improved by a large and diverse training set. In summary this work shows that AI based methodologies can be used for automated annotation and generation of interesting and frequent events. Such techniques provide an alternative means to human annotation.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d1575b36f7b8" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>