[Paper Review] Towards Discriminative Representation Learning for Unsupervised Person Re-identification

Takashi Isobe, Dong Li+4 more|arXiv (Cornell University)|Aug 7, 2021

Video Surveillance and Tracking Methods84 references4 citations

TL;DR

This paper proposes a unified framework for unsupervised person re-identification that improves discriminative feature learning through cluster-wise contrastive learning, progressive domain adaptation, and Fourier space augmentation. By iteratively refining pseudo-labels, gradually transferring knowledge from source to target, and enforcing class separability in the frequency domain, the method achieves state-of-the-art performance, surpassing MMT by 8.1% to 11.4% mAP across multiple benchmarks.

ABSTRACT

In this work, we address the problem of unsupervised domain adaptation for person re-ID where annotations are available for the source domain but not for target. Previous methods typically follow a two-stage optimization pipeline, where the network is first pre-trained on source and then fine-tuned on target with pseudo labels created by feature clustering. Such methods sustain two main limitations. (1) The label noise may hinder the learning of discriminative features for recognizing target classes. (2) The domain gap may hinder knowledge transferring from source to target. We propose three types of technical schemes to alleviate these issues. First, we propose a cluster-wise contrastive learning algorithm (CCL) by iterative optimization of feature learning and cluster refinery to learn noise-tolerant representations in the unsupervised manner. Second, we adopt a progressive domain adaptation (PDA) strategy to gradually mitigate the domain gap between source and target data. Third, we propose Fourier augmentation (FA) for further maximizing the class separability of re-ID models by imposing extra constraints in the Fourier space. We observe that these proposed schemes are capable of facilitating the learning of discriminative feature representations. Experiments demonstrate that our method consistently achieves notable improvements over the state-of-the-art unsupervised re-ID methods on multiple benchmarks, e.g., surpassing MMT largely by 8.1\%, 9.9\%, 11.4\% and 11.1\% mAP on the Market-to-Duke, Duke-to-Market, Market-to-MSMT and Duke-to-MSMT tasks, respectively.

Motivation & Objective

• To address the limitations of existing unsupervised domain adaptation methods in person re-identification, particularly label noise and poor knowledge transfer from source to target domains.
• To reduce noise in pseudo-labels generated from clustering on unlabeled target data through iterative optimization with contrastive learning.
• To improve knowledge transfer from the labeled source domain to the unlabeled target domain by replacing two-stage training with a progressive domain adaptation strategy.
• To enhance class separability of learned features by introducing additional constraints in the Fourier domain.
• To achieve state-of-the-art performance on multiple cross-domain person re-ID benchmarks without requiring manual annotations on the target domain.

Proposed method

• Introduces cluster-wise contrastive learning (CCL), where pseudo-labels from clustering are used as supervision in a contrastive learning objective, with a momentum-based moving-average encoder and a dynamic queue for negative samples.
• Employs a progressive domain adaptation (PDA) strategy that jointly trains on source and target domains with dynamically adjusted loss weights, gradually shifting from source to target supervision to improve knowledge transfer.
• Applies Fourier augmentation (FA) by computing additional loss functions on amplitude spectrum features, treating them as nonlinear transformations to enhance discriminative capacity.
• Uses a unified training pipeline that alternates between feature clustering and contrastive learning to refine pseudo-labels and improve representation quality iteratively.
• Implements a multi-branch head for joint training in both spatial and Fourier spaces, with learnable loss weights to balance the two supervisory signals.
• Employs a momentum encoder and a memory queue to stabilize training and provide diverse negative samples across mini-batches.

Figure 1: Given labeled source data and unlabeled target data, our goal is to learn feature representations for recognizing target classes. For this unsupervised domain adaptation task, we propose three technical schemes to learn discriminative target features: (a) cluster-wise contrastive learning,

Experimental results

Research questions

• RQ1 Can cluster-wise contrastive learning reduce label noise in pseudo-labeled unsupervised re-ID and improve feature discriminability?
• RQ2 Does progressive domain adaptation outperform standard two-stage fine-tuning in transferring knowledge from source to target domains?
• RQ3 Can Fourier space constraints enhance class separability and improve re-ID performance beyond spatial features alone?
• RQ4 How do different loss weight configurations and queue sizes affect the robustness and performance of the proposed framework?
• RQ5 To what extent does the combination of CCL, PDA, and FA lead to consistent improvements across diverse cross-domain re-ID benchmarks?

Key findings

• The proposed method achieves 69.4% mAP and 82.7% rank-1 on Market-to-Duke, surpassing the previous SOTA method MMT by 11.4% mAP.
• On Duke-to-Market, the method achieves 82.2% mAP and 93.6% rank-1, improving MMT by 11.1% mAP.
• The ablation study confirms that progressive domain adaptation with a 3-step policy outperforms static weight settings and two-stage training, with a 7.3% mAP gain over static 0.2/0.8 weights.
• Fourier augmentation combined with spatial features yields 69.4% mAP on Market-to-Duke, outperforming either space alone (67.3% or 67.6%) and showing consistent gains across all benchmarks.
• Optimal hyperparameters are found to be δ=0.1 and γ=0.7, which yield the best performance across all evaluated settings.
• The method demonstrates robustness to queue size and positive instance sampling, with |N_past|=1024 providing the best performance.

Figure 2: Illustration of the proposed unified framework for unsupervised cross-domain person re-ID.