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.

Domain Adaptive Person Re-Identification via Coupling Optimization

“…Comparison with recent works shows that our methods outperform state-of-theart works by a large margin. For instance, using DukeMTMC-reID as the source domain, our method achieves Rank-1 accuracy of 88.3% on Market-1501, outperforming recent PAST [49] and SSG [7] by 9.9% and 8.3%, respectively. We also test our method in unsupervised scenario, i.e., training the ReID model only with GLO.…”

“…However, local label prediction is not precise and will mislead the distance optimization. Based on assigned labels, some works adopt triplet loss for model training [7,33,46,49]. However, these optimization methods fail to consider the noise in predicted labels on unlabeled dataset.…”

Attribute Memory Transfer Network for Unsupervised Cross-Domain Person Re-Identification

“…For another aspect, several researches [8], [19], [30], [45] introduce a dependable pseudo identity labels generator to assist the training in unlabeled target domain. Specifically, the works [8], [30], [45] utilize the density-based clustering algorithm [3], [7] for label estimation, which has improved the pseudo-label reliability. In addition, the works [8], [45] also employ part division as the supplemental feature to obtain satisfactory performance.…”

“…Specifically, the works [8], [30], [45] utilize the density-based clustering algorithm [3], [7] for label estimation, which has improved the pseudo-label reliability. In addition, the works [8], [45] also employ part division as the supplemental feature to obtain satisfactory performance. In practical, these methods still obtain limited performance because they neglect the detailed local information in unlabeled target data.…”

“…A major solution for UCPR is attempting to transfer valuable information from a labeled source domain to a completely unlabeled target dataset [31], [34], where consists two major obstacles. Firstly, the pedestrian samples from different domains present a distribution shift [29], [36], [45]. Several models attempt to narrow down the pixel-level and feature-level domain gaps using Generative Adversarial Network (GAN) [31], [36] and Mean Maximum Discrepancy (MMD) [4], [33], separately.…”