[论文解读] Group-aware Label Transfer for Domain Adaptive Person Re-identification

Kecheng Zheng, Wu Liu+4 more|arXiv (Cornell University)|Mar 23, 2021

Video Surveillance and Tracking Methods参考文献 32被引用 17

一句话总结

本文提出了一种名为Group-aware Label Transfer（GLT）的新方法，用于无监督域自适应行人重识别（ReID），通过基于最优传输的标签迁移和一种群体感知策略，实现实时、相互优化的伪标签与表征学习。GLT在Duke→Market上达到92.2%的top-1准确率，在Market→Duke上达到82.0%，实现了当前最优性能，显著缩小了无监督与监督ReID之间的差距。

ABSTRACT

Unsupervised Domain Adaptive (UDA) person re-identification (ReID) aims at adapting the model trained on a labeled source-domain dataset to a target-domain dataset without any further annotations. Most successful UDA-ReID approaches combine clustering-based pseudo-label prediction with representation learning and perform the two steps in an alternating fashion. However, offline interaction between these two steps may allow noisy pseudo labels to substantially hinder the capability of the model. In this paper, we propose a Group-aware Label Transfer (GLT) algorithm, which enables the online interaction and mutual promotion of pseudo-label prediction and representation learning. Specifically, a label transfer algorithm simultaneously uses pseudo labels to train the data while refining the pseudo labels as an online clustering algorithm. It treats the online label refinery problem as an optimal transport problem, which explores the minimum cost for assigning M samples to N pseudo labels. More importantly, we introduce a group-aware strategy to assign implicit attribute group IDs to samples. The combination of the online label refining algorithm and the group-aware strategy can better correct the noisy pseudo label in an online fashion and narrow down the search space of the target identity. The effectiveness of the proposed GLT is demonstrated by the experimental results (Rank-1 accuracy) for Market1501$ o$DukeMTMC (82.0\%) and DukeMTMC$ o$Market1501 (92.2\%), remarkably closing the gap between unsupervised and supervised performance on person re-identification.

研究动机与目标

• 解决无监督域自适应（UDA）行人重识别（ReID）中伪标签噪声的问题，该问题会阻碍模型在不同域之间的泛化能力。
• 通过实现伪标签预测与表征学习之间的实时、在线交互，克服传统离线、迭代式伪标签标注与聚类方法的局限性。
• 通过引入基于隐式属性的群体ID，减少目标域身份匹配的搜索空间，实现群体感知聚类。
• 通过集成带有加权对比损失的目标实例记忆库，增强特征表示，提升模型鲁棒性与准确性。
• 通过在训练过程中动态优化伪标签，最小化监督与无监督ReID之间的性能差距。

提出的方法

• 将在线伪标签优化建模为最优传输问题，利用Sinkhorn-Knopp算法最小化将M个目标样本分配给N个伪标签的成本。
• 提出一种群体感知策略，为样本分配隐式属性群体ID，实现在无需人工标注的情况下进行多粒度聚类。
• 将标签迁移机制与群体感知聚类相结合，在训练过程中实时优化噪声伪标签，逐步提升标签质量。
• 设计一个带有加权对比损失的目标实例记忆库，通过挖掘全局正负样本增强特征表示。
• 整合三元组损失与加权对比损失，提升嵌入空间中的特征判别性与泛化能力。
• 采用统一目标端到端训练模型，联合优化表征学习与在线伪标签优化。

Figure 1: Illustration of conventional methods and our group-aware label transfer method. In our method, each instance is assigned to multiple prototypes with different granularity for generating multi-group pseudo labels, and then its noisy multi-group pseudo labels are online refined. By learning

实验结果

研究问题

• RQ1 在线、交互式伪标签优化在无监督行人ReID中的域自适应性能方面有何提升作用？
• RQ2 将群体感知聚类与隐式属性群体ID结合，能在多大程度上减少搜索空间并提升身份预测准确率？
• RQ3 基于最优传输的标签迁移是否在鲁棒性与收敛速度方面优于传统离线聚类方法？
• RQ4 在缺乏真实标签的情况下，记忆库与加权对比损失的结合如何增强特征表示？
• RQ5 群体感知标签与在线伪标签优化的联合影响是什么，能否有效缩小监督与无监督ReID之间的性能差距？

主要发现

• 在Duke→Market ReID基准上，GLT达到92.2%的top-1准确率，显著缩小了与监督性能（94.1% top-1）的差距。
• 在Market→Duke基准上，GLT达到82.0%的top-1准确率，展现出对域偏移的强大泛化能力。
• 消融实验表明，仅使用群体感知标签迁移即可使top-1准确率相比基线提升11.3个百分点，凸显其有效性。
• 标签迁移与目标实例记忆库结合加权对比损失后，Duke→Market上的top-1准确率从88.1%提升至89.0%。
• 同时引入三元组损失与加权对比损失后，Duke→Market的mAP提升3.2%，Market→Duke提升2.1%，表明二者具有互补优势。
• 若移除三元组损失或加权对比损失，mAP分别下降2.6%与2.8%，证实二者在性能中起关键作用。

Figure 2: Illustration of our group-aware label transfer framework (GLT), where the group-aware label transfer and the ReID model are alternately optimized to improve each other towards the final objective of accurate person re-ID. In the model pre-training stage, we pre-train the network using sour