Compression of Convolutional Neural Networks for Efficient Real-Time Person Re-Identification Applications

SPORTLOGiQ公司开发先进的体育分析软件来跟踪出现在体育游戏视频中的球员的位置和动作。这个项目的重点是在体育视频中的人物再识别任务，其中相同的球员应该在多个分布式摄像机视点中被识别，或者在单个摄像机中跨时间被识别。在这种情况下，由于玩家捕获条件、背景杂乱和遮挡的变化，人员重新识别算法的性能可能会下降。计算复杂性是实时应用程序的另一个重要考虑因素。该项目旨在为SPORTLOGiQ提供设计准确而高效的实时人员再识别系统所需的专业知识。鉴于深度学习架构在许多具有挑战性的视觉识别问题上实现了最先进的准确性，SPORTLOGiQ寻求开发和评估深度卷积神经网络（cnn），以准确地重新识别体育视频中的人物。然而，由于这些cnn代表了实时应用的复杂解决方案，该项目寻求开发专门的技术来压缩cnn，以减少它们的时间和内存复杂性。在这个项目中，将对文献中提出的几种技术的性能进行评估和比较，以提高cnn的计算效率，用于体育比赛视频的实时人物再识别。其中包括搜索缩减、特征选择和参数修剪的高级技术。特别是，该项目将主要关注有前途的新的过滤器级修剪技术，该技术可以根据从其层中提取的统计信息同时加速和压缩CNN。在这个项目中评估的技术对SPORTLOGiQ和计算机视觉和机器学习社区非常感兴趣。本研究项目的目的是解决体育视频分析背景下的计算机视觉基本问题之一。

SPORTLOGiQ Inc. develops advanced sports analytics software to track the location and actions of playersappearing in sports game videos. This project focuses on the task of person re-identification in sports videos,where the same players should be recognized across multiple distributed cameras viewpoint, or across timewithin a single camera. In this context, the performance of person re-identification algorithms can decline dueto variations in player capture conditions, background clutter, and occlusions. Computational complexity isanother important consideration for real-time applications.This project aims to provide SPORTLOGiQ with the expertise needed to design accurate yet efficient systemsfor real-time person re-identification. Given the state-of-the-art accuracy achieved with deep learningarchitectures on many challenging visual recognition problems, SPORTLOGiQ seeks to develop and evaluatedeep convolutional neural networks (CNNs) for accurate person re-identification in sports videos. However,since these CNNs represent complex solutions for real-time applications, this project seeks to developspecialized techniques for compression of CNNs, to reduce their time and memory complexity. In this project,the performance of several techniques proposed in the literature to increase the computational efficiency ofCNNs will be evaluated and compared for real-time person re-identification on sports game videos. Theseinclude advanced techniques for search reduction, features selection, and parameter pruning. In particular, thisproject will focus mostly on promising new filter-level pruning techniques that can simultaneously accelerateand compress a CNN based on statistical information extracted from its layers. The techniques evaluated in thisproject are of great interest to SPORTLOGiQ and to the computer vision and machine learning communities ingeneral. The aim of this is research project is to solve one of the fundamental computer vision problems in thecontext of sports video analytics.