[Paper Review] Collaborative Attention Network for Person Re-identification

Wenpeng Li, Yongli Sun+4 more|arXiv (Cornell University)|Nov 29, 2019

Video Surveillance and Tracking Methods22 references4 citations

TL;DR

This paper proposes a Collaborative Attention Network (CAN) for person re-identification that enhances feature representation by combining adjacent local feature slices from multi-scale global features through an end-to-end attention mechanism. By preserving spatial and scale-related relationships between neighboring local regions, CAN achieves state-of-the-art performance, outperforming prior methods by 2.4% in mAP and 0.7% in rank-1 on Market-1501.

ABSTRACT

Jointly utilizing global and local features to improve model accuracy is becoming a popular approach for the person re-identification (ReID) problem, because previous works using global features alone have very limited capacity at extracting discriminative local patterns in the obtained feature representation. Existing works that attempt to collect local patterns either explicitly slice the global feature into several local pieces in a handcrafted way, or apply the attention mechanism to implicitly infer the importance of different local regions. In this paper, we show that by explicitly learning the importance of small local parts and part combinations, we can further improve the final feature representation for Re-ID. Specifically, we first separate the global feature into multiple local slices at different scale with a proposed multi-branch structure. Then we introduce the Collaborative Attention Network (CAN) to automatically learn the combination of features from adjacent slices. In this way, the combination keeps the intrinsic relation between adjacent features across local regions and scales, without losing information by partitioning the global features. Experiment results on several widely-used public datasets including Market-1501, DukeMTMC-ReID and CUHK03 prove that the proposed method outperforms many existing state-of-the-art methods.

Motivation & Objective

• Address the limitation of global-only features in capturing discriminative local patterns for person re-identification.
• Overcome the information loss and arbitrary part definition in handcrafted local feature slicing methods.
• Improve part-based feature learning by modeling relationships between adjacent local regions across scales.
• Develop an end-to-end trainable mechanism that dynamically attends to collaborative local features without partitioning global features.
• Achieve superior performance on benchmark datasets by combining global and locally enhanced representations through a multi-branch architecture and joint optimization.

Proposed method

• Split the global feature map into multiple multi-scale local slices using a multi-branch network structure.
• Introduce a Collaborative Attention Network (CAN) that fuses features from adjacent local slices to preserve spatial and scale-related relationships.
• Use learnable attention weights to dynamically emphasize informative neighboring regions across local parts.
• Integrate both global and local branches into triplet loss and center loss for joint optimization, enhancing discriminative capacity.
• Apply feature and weight normalization with cosine distance for similarity measurement during inference.
• Train the model using Adam optimizer with a cosine decay learning rate schedule and multi-loss function (CrossEntropy, Triplet, Center Loss).

Experimental results

Research questions

• RQ1 Can combining adjacent local feature slices improve discriminative representation in person re-identification beyond fixed-part or global-only methods?
• RQ2 Does modeling inter-slice relationships across scales and spatial neighborhoods lead to better performance than independent local feature extraction?
• RQ3 What is the optimal number of local slices that maximizes performance without causing overfitting or information loss?
• RQ4 Can an end-to-end collaborative attention mechanism outperform hard-attention or fixed-part fusion strategies in part-based Re-ID?
• RQ5 Does joint optimization of global and local branches via triplet and center loss enhance feature discrimination on standard benchmarks?

Key findings

• The proposed CAN achieves 90.6% mAP and 96.4% rank-1 on Market-1501, surpassing the previous SOTA (Pyramid: 88.2% mAP, 95.7% rank-1) by 2.4 and 0.7 percentage points respectively.
• On DukeMTMC-ReID, CAN achieves 82.3% mAP and 91.6% rank-1, outperforming the previous best method (Pyramid) by a significant margin.
• On CUHK03, CAN achieves 82.8% mAP and 86.1% rank-1, exceeding all prior published methods, including those tested on both labeled and detected images.
• Experiments show that increasing the number of local slices beyond an optimal point reduces accuracy, indicating a trade-off between granularity and performance.
• The ablation study confirms that collaborative attention between adjacent slices improves feature representation more effectively than independent or fixed-part fusion.
• Joint optimization of global and local branches via triplet and center loss significantly enhances model generalization and discriminative power.