Multi-scale feature interaction pseudo-label unsupervised domain adaptation for person re-identification

Person re-identification (ReID) is a computer vision task that matches target persons from images or video sequences and has wide application prospects. Although supervised ReID methods have achieved excellent performance on single-domain datasets by introducing deep learning technologies, they are difficult to apply directly to real scenarios because of the time-consuming and labor-intensive process of collecting and annotating pedestrian images. Furthermore, domain gaps between different datasets significantly degrade the recognition performance of these models. Due to these challenges, unsupervised domain adaptation (UDA) methods have been introduced to the ReID field to reduce the dependency on labeled datasets and enhance generalizability across domains. UDA ReID aims to leverage labeled source domain data to improve model performance on unlabeled target domains. Currently, there are roughly two ways to tackle the problem: generative-based and pseudo-label clustering-based methods. Generative methods mainly use generative adversarial networks to construct new images that bridge domain gaps between datasets or cameras. However, the unstable training process of generative adversarial networks often results in unrealistic or noisy images. In contrast, clustering-based methods employ clustering algorithms or strategies to generate pseudo-labels for target domain data, enabling supervised training. These methods are relatively straightforward in design and achieve outstanding performance among various UDA approaches. Nevertheless, existing clustering-based ReID methods face challenges such as limited receptive fields and weak connections between global and local features, leading to suboptimal feature extraction and recognition accuracy. To address this issue, we propose a multi-scale feature interaction method, which can fully extract features and effectively improve the recognition accuracy. First, the feature squeeze attention mechanism compresses input features along horizontal and vertical axes, enhancing local detail extraction while aggregating global features. Second, the residual feature interaction module expands the receptive field through multi-scale interactions and encodes global contextual information into features, strengthening connections between global and local features. Third, the partial convolution bottleneck module selectively applies convolution operations to a subset of input channels, reducing redundant computations while preserving spatial feature information. The method proposed in this paper was validated on three public datasets, and the results showed that the performance has been significantly improved compared to the baseline method. This method achieved the mAP (mean average precision) of 82.9%, Rank-1 accuracy of 93.7% on the Market-1501 dataset, mAP of 68.7%, and Rank-1 accuracy of 82.7% on the DukeMTMC-reID dataset, and the mAP of 26.6% and Rank-1 accuracy of 54.7% on the MSMT17 dataset. Ablation studies further confirm the independent contributions of various modules and the significant enhancements in feature extraction and computational efficiency achieved by their integration. These results demonstrate its superiority by extracting comprehensive multi-scale features and improving cross-domain recognition accuracy.