TROPE: Triplet-Guided Feature Refinement for Person Re-Identification

Person Re-Identification (PRid) has garnered research attention with the rising popularity of intelligent video surveillance. Deep learning methods like Convolutional Neural Networks (CNN) are vital in PRid. The CNNs acquire image characteristics that are distinctive, referred to as image features, by analyzing the entire image. This process enables them to recognize and differentiate between various images. However, in the case of PRid, this search across the entire image may lead the model to emphasize distinctive image features of the background while neglecting subtle but essential distinguishing regions of the person. This tendency can be observed in the heat maps generated from trained models. Therefore, it is crucial to direct the model's attention towards the vital regions of a person. Relying solely on global features might be limited in effectiveness, as it does not sufficiently capture essential finer details. Therefore, it becomes necessary to pinpoint significant features at a local level and guide the model to prioritize these features for improved results. Inspired to identify and prioritize vital regions based on local features, we propose TROPE (Triplet-Guided Feature Refinement for Person Re-Identification) in this paper. The technique involves analyzing the intermediate features of hard positive and hard negative images and generating weight vectors. These vectors are utilized to shift the attention of the network to specific regions of interest. The proposed method achieves 87.5% mAP and 95.6% Rank1 accuracy on Market1501. 77.6% mAP and 88.8% Rank1 accuracy on DukeMTMC and 78.5% mAP, 80.1% Rank1 accuracy on CUHK03 dataset.