A Filtering Non-Discriminative Features Method for Person Re-Identification Based on Transformer

Due to the extensive use of convolution and down-sampling modules in CNN architectures, models based on this structure inevitably suffer from image information loss during computation, adversely affecting the performance of corresponding computer vision tasks. This paper investigates some semi-supervised person ReID (re-identification)
^me
thods, and proposes a model fine-tuning approach named FNDF-T-ReID for filtering non-discriminative features based on a pre-trained transformer ReID model. The method of semantic part classification in the self-supervised training phase combines ratio filtering and K-means clustering algorithm, making the training process more robust and flexible. In the supervised training phase, the model training tokens incorporate the camera labels of each image, mitigating the impact of image style differences from different cameras on the identification of pedestrians in the images. To further enhance the model's adaptability to target domain images, dataset extension methods are used to optimize the model training process. Finally, the model is trained on multiple person ReID datasets and evaluated using metrics of mAP and Rank-1, Experimental results demonstrate that the proposed method maintains high stability during training and outperforms baseline models on several metrics, achieving state-of-the-art performance.