[Paper Review] Parameter-Free Spatial Attention Network for Person Re-Identification

Haoran Wang, Yue Fan+3 more|arXiv (Cornell University)|Nov 29, 2018

Video Surveillance and Tracking Methods21 references78 citations

TL;DR

The paper introduces a parameter-free spatial attention layer added before GAP to model spatial relations, improving GAP-based Re-ID models and achieving state-of-the-art results on four benchmarks.

ABSTRACT

Global average pooling (GAP) allows to localize discriminative information for recognition [40]. While GAP helps the convolution neural network to attend to the most discriminative features of an object, it may suffer if that information is missing e.g. due to camera viewpoint changes. To circumvent this issue, we argue that it is advantageous to attend to the global configuration of the object by modeling spatial relations among high-level features. We propose a novel architecture for Person Re-Identification, based on a novel parameter-free spatial attention layer introducing spatial relations among the feature map activations back to the model. Our spatial attention layer consistently improves the performance over the model without it. Results on four benchmarks demonstrate a superiority of our model over the state-of-the-art achieving rank-1 accuracy of 94.7% on Market-1501, 89.0% on DukeMTMC-ReID, 74.9% on CUHK03-labeled and 69.7% on CUHK03-detected.

Motivation & Objective

• Address the limitations of Global Average Pooling (GAP) in capturing spatial relations for fine-grained person Re-ID.
• Propose a parameter-free spatial attention layer that reweights spatial positions based on activations across channels.
• Integrate the spatial attention layer into a GAP-based classifier with deep supervision and part-based features.
• Show that the proposed approach yields robust and distributed attention, improving cross-view Re-ID performance.

Proposed method

• Introduce a parameter-free spatial attention layer that sums activations across channels per location, applies a softmax over spatial positions, and reweights the feature map accordingly.
• Insert the spatial attention layer before GAP; the layer is proportional to the total intensity at each spatial position and does not introduce trainable parameters.
• Form the overall architecture by combining ResNet-50 backbone, deep supervision branches, and PCB-style part classifiers, with the spatial attention layer applied in auxiliary branches.
• Train with a composite loss L_total = (1-λ) sum of auxiliary losses + λ sum of deeply supervised losses + λ global loss, using six part-level features plus global features.

Experimental results

Research questions

• RQ1 Can a parameter-free spatial attention mechanism improve robustness of GAP-based features for person Re-ID under viewpoint changes and occlusions?
• RQ2 How does distributing attention across spatial locations affect class activation maps and gradient flow in Re-ID models?
• RQ3 Does integrating spatial attention with deep supervision and part-based features yield state-of-the-art performance on standard Re-ID benchmarks?

Key findings

• The spatial attention layer consistently improves performance over GAP without attention across experiments.
• On Market-1501, the method achieves 94.7% Rank-1 and 91.7% mAP in the reported table.
• On DukeMTMC-ReID, it reaches 89.0% Rank-1 and 85.9% mAP.
• On CUHK03 (labeled), it reaches 74.9% Rank-1 and 76.5% mAP.
• On CUHK03 (detected), it reaches 69.7% Rank-1 and 72.2% mAP.
• Qualitative CAM visualizations show more distributed attention and robustness to partial occlusions; ablations confirm the gains from SA (SA1/SA2) and architecture choices.