Source-linked AI summary
Self-training with progressive augmentation for unsupervised cross-domain person re-identification
Xinyu Zhang, Jiewei Cao, Chunhua Shen, Mingyu You
TL;DR
The paper addresses transferring person Re-ID models from labelled source domains to unlabelled target domains, where domain shift and unreliable pseudo labels hinder adaptation. It proposes PAST, which alternates conservative local-structure learning with promoting global-distribution learning and introduces a label-free ranking-based triplet loss. Experiments report that PAST outperforms state-of-the-art unsupervised cross-domain algorithms by a large margin.
Problem
Unsupervised cross-domain person Re-ID must transfer learned information from a labelled source domain to an unlabelled target domain despite domain shift and unreliable pseudo labels.
Method
PAST alternates conservative triplet-based learning and promoting classification-based learning, using ranking-based triplet loss to reduce reliance on unreliable pseudo labels.
Results
PAST outperforms state-of-the-art unsupervised cross-domain algorithms by a large margin.
Takeaways & Limitations
Alternating local- and global-information learning progressively improves model generalization and feature representation on unlabelled target domains.
Takeaways & Limitations
The authors identify extending PAST to other unsupervised cross-domain applications, such as face recognition and image retrieval, as future work.
Abstract
from arXiv · showhide
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
1. Introduction
Unsupervised cross-domain person Re-ID must transfer knowledge from labelled source data to unlabelled target data despite domain shift and unreliable pseudo labels. PAST alternates conservative and promoting stages to progressively improve target-domain generalization.
- Research challenge: Unsupervised cross-domain Re-ID transfers a source-trained model to an unlabelled target domain with inconsistent characteristics.Direct source-to-target application performs unsatisfactorily because of domain shift.
- Research challenge: Pseudo-label-based methods depend strongly on clustering quality, while embedding and classification losses have complementary weaknesses.Embedding losses can converge slowly or reach suboptimal results, whereas classification losses depend heavily on pseudo-label quality.
- PAST framework: PAST alternates conservative and promoting stages to restrain early error amplification and progressively incorporate more confidently labelled examples.The conservative stage uses local information, while the promoting stage uses global data distribution information.
- Conservative stage: Ranking-based triplet loss selects positive and negative examples from ranked similarity groups without relying solely on pseudo labels.For each anchor, positives come from the top η ranked images and negatives from ranks (η, 2η].
- Promoting stage: The promoting stage treats clusters as classes and applies softmax cross-entropy to exploit global distribution and improve inter-class separability.This complements triplet-based local-structure learning, which can become unstable or settle at a local optimum.
- Experimental outcome: Experiments on three large-scale datasets indicate that PAST is effective for unsupervised cross-domain person Re-ID.The conclusion reports that PAST outperforms state-of-the-art unsupervised cross-domain algorithms by a large margin.
2. Our Method
PAST alternates conservative triplet-based training with promoting classification-based training to refine target-domain representations progressively. It uses ranking and clustering information to capture local structure, then global cluster structure, while reinitializing the classifier from current cluster features.
- Overall Framework: PAST alternates conservative and promoting stages to optimize the target-domain model progressively.The framework uses triplet-based losses in the conservative stage and classification loss in the promoting stage.
- Conservative Stage: The conservative stage extracts target features, computes ranking-based distances, clusters images with HDBSCAN, and retains clustered images as the updated training set.Images without cluster assignments are discarded from further optimization.
- Conservative Stage: CTL and RTL jointly capture local data structure, with RTL selecting positives and negatives by ranking positions rather than relying on pseudo labels.RTL samples positives from the top η neighbors and negatives from ranks (η, 2η], using a soft margin based on relative ranking positions.
- Promoting Stage: The promoting stage treats clusters as classes and applies softmax cross-entropy to exploit the global distribution of training samples.A fully connected classifier is added after the triplet-based update and initialized from the current training-set features.
- Promoting Stage: The classifier is reinitialized from each cluster’s mean feature, using Wc = Fc, to avoid random-initialization fluctuations during iterative training.The classifier must be initialized whenever HDBSCAN changes the number of clusters.
- Training Procedure: The full procedure repeatedly extracts features, computes rankings, updates HDBSCAN clusters, trains with CTL and RTL, extracts updated features, and optimizes the classifier.This alternating process is implemented as the PAST self-training algorithm.
3. Experiments
Experiments evaluate PAST on three large-scale person Re-ID datasets, test its components and clustering choices, and compare it with unsupervised cross-domain methods. The results show benefits from combining conservative and promoting stages, with HDBSCAN and suitable parameter settings further improving performance.
- Experimental Setup: Experiments use Market-1501, DukeMTMC-Re-ID, and CUHK03 for unsupervised cross-domain person Re-ID evaluation.The implementation uses PCB with a ResNet-50 backbone, nine feature regions, 256-dimensional embeddings, and cosine-distance ranking evaluated by CMC and mAP.
- Ablation Study: 18.49% and 12.14% Rank-1 gains over k-reciprocal encoding are obtained by CTL on M→D and D→M, respectively.Using only RTL increases Rank-1 and mAP by 21% and 12.64% on M→D, and 12.91% and 5.69% on D→M; combining CTL and RTL performs better on both tasks.
- Clustering Comparison: HDBSCAN achieves 54.26% mAP and 72.35% Rank-1 on M→D with PAST, exceeding k-means by 4.29% and 3.41%.It also exceeds DBSCAN by 1.19% mAP and 0.45% Rank-1, while PAST outperforms the conservative stage across clustering methods.
- State-of-the-Art Comparison: PAST improves over the conservative stage by 4.71% Rank-1 and 5.21% mAP on C→D and outperforms competing cross-domain methods in reported comparisons.The conservative stage alone is already competitive, including gains over EANet on M→D.
- Parameter Analysis: λ=0.5 gives the best D→M result, while Smin=10 yields superior accuracy and 625 pseudo identities, closest to the true 751.Large or small λ limits improvement; changing Smin produces substantial variance in the number of pseudo identities.
4. Conclusion
PAST alternates conservative and promoting stages to combine local and global information for unsupervised cross-domain person re-identification. The authors report improved target-domain generalization and performance exceeding state-of-the-art algorithms, while identifying extension to other applications as future work.
- PAST alternates conservative and promoting stages to provide complementary local and global information.The conservative stage uses triplet-based losses, while the promoting stage extracts global information.
- The ranking-based triplet loss reduces reliance on clustering quality by providing a label-free objective for selecting triplet samples.
- Extensive experiments show that PAST outperforms state-of-the-art unsupervised cross-domain algorithms by a large margin.
- The authors plan to extend PAST to unsupervised cross-domain face recognition and image retrieval.
1. More Experimental Results
Additional experiments examine ranking-based triplet selection, feature representations, and clustering behavior. The reported analyses indicate robust ranking-position performance, improved clustering assignments, and more reliable triplets during training.
- Ranking-based triplet selection: η values of 20 or larger produce nearly identical and competitive results for Duke-to-Market transfer, so the experiments use η = 20.
- Feature representation: PAST embeddings gather images from the same identity and separate images from different classes in t-SNE visualizations.
- Triplet selection and pseudo-label quality: PAST progressively improves clustering assignments and pseudo-label reliability across training iterations.More reliable pseudo-labels can then be used in the promoting stage to further improve model generalization.
- Triplet selection and pseudo-label quality: RTL can select correct positive samples from a noisy cluster because it relies on similarity rankings rather than cluster assignments alone.