[论文解读] Robust Cross-View Gait Recognition with Evidence: A Discriminant Gait GAN (DiGGAN) Approach

Bingzhang Hu, Yu Guan+4 more|arXiv (Cornell University)|Nov 26, 2018

Gait Recognition and Analysis参考文献 45被引用 16

一句话总结

该论文提出DiGGAN，一种基于判别式生成对抗网络（GAN）的框架，用于实现鲁棒的跨视角步态识别，能够生成视角不变的特征，并在目标视角下合成步态能量图像（GEI）作为可解释的证据。该模型在OU-MVLP和CASIA-B数据集上优于当前最先进方法，识别错误率降低至1.2%（rank-1），验证任务的等错误率（EER）为2.4%，同时实现了稳定且人类可解释的证据生成，支持决策验证。

ABSTRACT

Gait as a biometric trait has attracted much attention in many security and privacy applications such as identity recognition and authentication, during the last few decades. Because of its nature as a long-distance biometric trait, gait can be easily collected and used to identify individuals non-intrusively through CCTV cameras. However, it is very difficult to develop robust automated gait recognition systems, since gait may be affected by many covariate factors such as clothing, walking speed, camera view angle etc. Out of them, large view angle changes has been deemed as the most challenging factor as it can alter the overall gait appearance substantially. Existing works on gait recognition are far from enough to provide satisfying performances because of such view changes. Furthermore, very few works have considered evidences -- the demonstrable information revealing the reliabilities of decisions, which are regarded as important demands in machine learning-based recognition/authentication applications. To address these issues, in this paper we propose a Discriminant Gait Generative Adversarial Network, namely DiGGAN, which can effectively extract view-invariant features for cross-view gait recognition; and more importantly, to transfer gait images to different views -- serving as evidences and showing how the decisions have been made. Quantitative experiments have been conducted on the two most popular cross-view gait datasets, the OU-MVLP and CASIA-B, where the proposed DiGGAN has outperformed state-of-the-art methods. Qualitative analysis has also been provided and demonstrates the proposed DiGGAN's capability in providing evidences.

研究动机与目标

• 为解决步态识别中大视角变化带来的挑战，此类变化会显著降低识别性能。
• 提出一种框架，通过生成不同视角下的合成GEI作为可解释证据，提升安全与隐私应用中决策的透明度。
• 通过引入局部增强模块提升特征的判别能力，增强跨视角的泛化性能。
• 实现任意视角到任意视角的步态图像翻译，确保证据的一致性与跨视角比较的可行性。
• 通过引入基于证据的推理机制，弥合自动化步态识别与实际部署之间的差距。

提出的方法

• 设计了一种条件GAN框架，通过解耦身份与视角因素，学习视角不变的步态表征。
• 生成器网络以源GEI和目标视角角度作为输入，利用视角不变特征提取器与视角条件模块，合成指定视角下的GEI。
• 引入局部增强模块以优化视角不变特征，提升其判别能力与鲁棒性。
• 采用对抗损失、重建损失以及对比损失或三元组损失联合训练模型，确保特征紧凑性与身份一致性。
• 从生成器的瓶颈层提取视角不变特征，支持跨视角匹配与证据生成。
• 该框架支持任意视角之间的图像翻译，可从任意输入视角生成所有训练集中存在的视角角度的GEI。

实验结果

研究问题

• RQ1 基于GAN的模型能否学习到在大视角变化下仍具泛化能力的视角不变步态特征？
• RQ2 在不同视角下合成的步态图像（GEI）能否作为决策验证中可靠且可解释的证据？
• RQ3 在匹配步态样本时，特征层面的相似性与证据层面的相似性是否具有一致性？
• RQ4 引入局部增强模块是否能提升所学习步态表征的判别质量？
• RQ5 该模型是否能在跨数据集评估中保持性能，表明其在训练数据之外的鲁棒性？

主要发现

• 在OU-MVLP数据集上，DiGGAN实现了1.2%的rank-1识别错误率，优于以往最先进方法。
• 在OU-MVLP数据集的验证任务中，DiGGAN实现了2.4%的等错误率（EER），表现优于现有方法。
• 在CASIA-B数据集上，DiGGAN在对比损失设置下实现了1.3%的rank-1错误率与2.6%的EER，展现出强大的泛化能力。
• 定性分析表明，生成的GEI在不同视角下均保持了关键步态特征，如肢体与颈部姿态。
• 特征层面相似性与证据层面相似性之间的一致性较高，验证了生成GEI作为决策支持证据的可靠性。
• 模型成功将视角信息从身份特征中解耦，表现为在相同视角下，不同输入生成的GEI方差极低。