Self-training with progressive augmentation for unsupervised cross-domain person re-identification

Abstract: Person re-identification (Re-ID) has achieved great improvement with deep learning and a large amount of labelled training data. However, it remains a challenging task for adapting a model trained in a source domain of labelled data to a target domain of only unlabelled data available. In this work, we develop a self-training method with progressive augmentation framework (PAST) to promote the model performance progressively on the target dataset. Specially, our PAST framework consists of two stages, namely, conservative stage and promoting stage. The conservative stage captures the local structure of target-domain data points with triplet-based loss functions, leading to improved feature representations. The promoting stage continuously optimizes the network by appending a changeable classification layer to the last layer of the model, enabling the use of global information about the data distribution. Importantly, we propose a new self-training strategy that progressively augments the model capability by adopting conservative and promoting stages alternately. Furthermore, to improve the reliability of selected triplet samples, we introduce a ranking-based triplet loss in the conservative stage, which is a label-free objective function basing on the similarities between data pairs. Experiments demonstrate that the proposed method achieves state-of-the-art person Re-ID performance under the unsupervised cross-domain setting. Code is available at: https://tinyurl.com/PASTReID

• The paper presents a progressive framework (PAST) that alternates conservative and promoting stages to tackle domain shift in person re-ID.
• It employs clustering-based and ranking-based triplet losses in the conservative stage and softmax cross-entropy in the promoting stage for balanced feature learning.
• Experiments on Market-1501, DukeMTMC-Re-ID, and CUHK03 demonstrate significant improvements in Rank-1 accuracy and mAP compared to state-of-the-art methods.

Analysis of the PAST Framework for Unsupervised Cross-Domain Person Re-Identification

The paper in focus introduces a self-training method with a progressive augmentation framework (PAST) designed to enhance model performance for unsupervised cross-domain person re-identification (Re-ID). Within this domain, the challenge primarily revolves around transferring knowledge from a labeled source domain to an unlabeled target domain effectively, with the ultimate goal of overcoming the domain shift problem.

Key Innovations and Methodology

At the core of the proposed framework are two alternating training stages: the conservative stage and the promoting stage, which complement each other by iteratively refining model performance across domains.

1. Conservative Stage:
   • This phase employs a combination of clustering-based triplet loss (CTL) and an innovative ranking-based triplet loss (RTL).
   • The CTL uses pseudo labels generated from clusters to fine-tune the model through triplet selection within these clusters.
   • The RTL, a label-free loss function, constructs triplets by leveraging the similarity ranking of data points, offering a robust alternative to directly relying on potentially noisy pseudo labels.
2. Promoting Stage:
   • Once the model has been refined through local triplet selection, it undergoes further optimization using a softmax cross-entropy loss at this stage.
   • This leverages global data distribution, which allows for enhanced discriminative feature learning that significantly boosts model stability and generalization on the target dataset.

By alternating these stages, the PAST framework progressively improves model performance, effectively capturing both local and global data structures.

Experimental Results and Evaluation

The paper benchmarked its proposed framework on multiple large-scale datasets, including Market-1501, DukeMTMC-Re-ID, and CUHK03. The empirical results showed that PAST outperformed state-of-the-art unsupervised cross-domain Re-ID methods, with notable improvements in both Rank-1 accuracy and mAP metrics across different experimental settings (e.g., Market-1501 to DukeMTMC-Re-ID, and vice versa).

Implications and Future Directions

The PAST framework presents a significant advancement in the field of unsupervised domain adaptation in person Re-ID without requiring auxiliary domain-specific information such as camera IDs or pose annotations. This marks a step toward more robust and generalizable Re-ID systems applicable in real-world scenarios where labeled data is scarce.

However, as the framework relies on pseudo-label refinement and progressive training, further exploration could focus on automating and enhancing clustering techniques to refine pseudo-label quality. Additionally, integrating self-supervised learning strategies, which are gaining traction within AI communities, could complement this approach, providing richer feature representations and further minimizing reliance on external supervision.

This research adequately showcases the potential of harnessing progressive methods to overcome domain shifts, a crucial consideration for extending person Re-ID applications beyond specifically controlled environments. Future research could also investigate the applicability of the PAST framework leveraged for other domain adaptation tasks within computer vision, such as object detection or semantic segmentation, to further validate its generality and robustness.