[Paper Review] Enhancing Diversity in Teacher-Student Networks via Asymmetric branches for Unsupervised Person Re-identification

Hao Chen, Benoit Lagadec+1 more|arXiv (Cornell University)|Nov 27, 2020

Video Surveillance and Tracking Methods38 references39 citations

TL;DR

Introduces asymmetric branches inside Mean Teacher for unsupervised Re-ID to increase appearance- signature diversity and decouple teacher-student weights, improving unsupervised domain adaptation and fully unsupervised Re-ID.

ABSTRACT

The objective of unsupervised person re-identification (Re-ID) is to learn discriminative features without labor-intensive identity annotations. State-of-the-art unsupervised Re-ID methods assign pseudo labels to unlabeled images in the target domain and learn from these noisy pseudo labels. Recently introduced Mean Teacher Model is a promising way to mitigate the label noise. However, during the training, self-ensembled teacher-student networks quickly converge to a consensus which leads to a local minimum. We explore the possibility of using an asymmetric structure inside neural network to address this problem. First, asymmetric branches are proposed to extract features in different manners, which enhances the feature diversity in appearance signatures. Then, our proposed cross-branch supervision allows one branch to get supervision from the other branch, which transfers distinct knowledge and enhances the weight diversity between teacher and student networks. Extensive experiments show that our proposed method can significantly surpass the performance of previous work on both unsupervised domain adaptation and fully unsupervised Re-ID tasks.

Motivation & Objective

• Motivate reducing pseudo-label noise in unsupervised Re-ID by increasing feature and weight diversity.
• Propose asymmetric branches to diversify appearance signatures.
• Introduce cross-branch supervision to prevent coupling of teacher-student networks.
• Demonstrate improvements on unsupervised domain adaptation and fully unsupervised Re-ID tasks.

Proposed method

• Duplicate the last layers of a ResNet backbone to create two asymmetric branches with different depths and pooling (GAP vs GMP).
• Concatenate features from both branches to form richer appearance signatures for clustering-based pseudo-labels.
• Use cross-branch teacher-student supervision where each branch’s teacher guides the opposite branch’s student via cross-entropy and soft triplet losses.
• Compute dynamic classifiers per cluster at each epoch to handle varying cluster counts in DBSCAN-based pseudo-labeling.
• Train with Mean Teacher where teacher weights are EMA of the student, augmented by cross-branch soft labels and soft triplet constraints.

Experimental results

Research questions

• RQ1 Can heterogeneous (asymmetric) branches within a Mean Teacher framework sustain diversity between teacher and student to avoid coupling?
• RQ2 Do cross-branch supervision signals improve the quality of soft pseudo-labels and discriminative power of features for Re-ID?
• RQ3 How does ABMT perform compared with state-of-the-art unsupervised domain adaptation and fully unsupervised Re-ID methods on standard benchmarks?

Key findings

• Asymmetric branches substantially improve performance during target domain adaptation.
• Dynamic clustering with DBSCAN plus cross-branch supervision yields better hard and soft pseudo-labels.
• ABMT outperforms previous Mean-Teacher based UDA methods on Market→Duke and Duke→Market tasks (e.g., 69.1% mAP and 82.0% Rank-1 on Market→Duke).
• Cross-branch supervision maintains higher teacher-student feature diversity throughout training, reducing coupling.
• Backbone enhancements (e.g., IBN-ResNet50) further boost results on all tasks.