Prototypical contrastive learning-based transformer for occluded person re-identification

This paper introduces PCL-Former, a Transformer model built on Prototypical Contrastive Learning to tackle the Occluded Person Re-Identification problem. Although pedestrian re-identification has seen significant progress in recent years, practical applications still face challenges. Pedestrians are often obstructed by vehicles, other individuals, or obstacles, making it difficult for traditional recognition methods to accurately extract useful features. To solve this problem, PCL-Former divides the image into multiple local segments by Vision Transformer and generates multiple class markers for each image, which represent different classes of features. In order to enhance the differentiation between class markers, this paper introduces the Prototypical Contrastive Learning loss function, which clusters the embeddings of the same class markers together and separates the embeddings of different class markers as much as possible. This method effectively solves the feature confusion problem in occlusion scenarios, enabling the model to better distinguish similar classes and improve recognition accuracy. We conducted thorough experimental validation of the method across multiple occluded person re-identification datasets. The results show that PCL-Former surpasses existing methods in occluded scenes, especially on the Occluded-Duke dataset. Compared to the baseline model, the mAP score and Rank-1 accuracy improved by 2.9% and 1.6%, respectively. Additionally, the model exhibits strong domain generalization capabilities and maintains high performance across different datasets.