[Paper Review] ABD-Net: Attentive but Diverse Person Re-Identification

Tianlong Chen, Shaojin Ding+6 more|ArXiv.org|Aug 3, 2019

Video Surveillance and Tracking Methods64 references64 citations

TL;DR

ABD-Net combines channel and position attention with a novel SVDO orthogonality regularizer to enforce diversity in both activations and weights, achieving state-of-the-art results on major Re-ID benchmarks.

ABSTRACT

Attention mechanism has been shown to be effective for person re-identification (Re-ID). However, the learned attentive feature embeddings which are often not naturally diverse nor uncorrelated, will compromise the retrieval performance based on the Euclidean distance. We advocate that enforcing diversity could greatly complement the power of attention. To this end, we propose an Attentive but Diverse Network (ABD-Net), which seamlessly integrates attention modules and diversity regularization throughout the entire network, to learn features that are representative, robust, and more discriminative. Specifically, we introduce a pair of complementary attention modules, focusing on channel aggregation and position awareness, respectively. Furthermore, a new efficient form of orthogonality constraint is derived to enforce orthogonality on both hidden activations and weights. Through careful ablation studies, we verify that the proposed attentive and diverse terms each contributes to the performance gains of ABD-Net. On three popular benchmarks, ABD-Net consistently outperforms existing state-of-the-art methods.

Motivation & Objective

• Motivate the integration of attention with feature diversity to improve Re-ID discriminability.
• Develop complementary attention modules (channel and position) to capture robust, discriminative features.
• Introduce SVDO-based orthogonality to reduce feature and weight correlations without heavy computation.
• Demonstrate that attention plus diversity yields better performance than attention alone or diversity alone.

Proposed method

• Implement a compound attention framework with Channel Attention Module (CAM) and Position Attention Module (PAM) to capture channel-wise and spatial dependencies.
• Apply a novel spectral value difference orthogonality (SVDO) regularization to both hidden activations (O.F.) and weights (O.W.) to promote diversity.
• Use a ResNet-50 backbone with dual-branch architecture (attentive and global branches) and concatenate their outputs for final embedding.
• Train with a loss combining cross-entropy, hard mining triplet loss, and SVDO penalties (L = Lxent + β_tr L_triplet + β_O.F L_O.F + β_O.W L_O.W).
• Approximate SVDO via power iteration to estimate top eigenvalues of FF^T, avoiding full SVD.

Experimental results

Research questions

• RQ1 Can combining channel and position attention improve robustness to misalignment and background perturbations in Re-ID?
• RQ2 Does enforcing diversity with SVDO reduce feature correlations and improve retrieval when using Euclidean distance?
• RQ3 How does ABD-Net perform relative to state-of-the-art on Market-1501, DukeMTMC-Re-ID, and MSMT17?
• RQ4 What is the contribution of attention versus diversity to overall gains in Re-ID performance?

Key findings

• ABD-Net consistently outperforms prior methods on three major Re-ID benchmarks.
• Joint use of CAM and PAM yields better performance than either attention module alone.
• SVDO-based regularization on features and weights reduces feature correlations and improves discriminability.
• Combining attention with diversity yields additional gains beyond attention or diversity alone.
• On Market-1501, ABD-Net achieves state-of-the-art results with significant gains when compared to attention-based baselines.
• On DukeMTMC-Re-ID and MSMT17, ABD-Net attains leading top-1 and mAP scores.