Multi-task loss design of person re-identification model
Person re-identification is a task that utilizes computer vision technology to discern the pres-ence of specific pedestrians within images.In tackling the problem of Re-ID models struggling to effec-tively learn similar local appearance between different pedestrians when employing identity labels,a ap-proach based on multi-task loss has been introduced.Initially,global and local features are extracted via the backbone network,with a pose estimation algorithm utilized to detect pedestrian body parts.Then,integrate the features of body parts with local features to form human pose-guided features.Sub-sequently,through a specially designed multi-task loss methodology,the model is guided to optimize both human pose-guided features and global features,thereby fortifying its robustness against occlusion and non-discriminative local appearances.The results indicate that this approach achieves precision rates of 59.7％/67.9％,88.4％/94.9％,and 80.6％/89.9％for mAP/Rank-1 across the Occluded-Duke,Market 1501,and DukeMTMC-reID datasets,respectively.To mitigate the impact of distribution discrepancies between training and testing datasets on the performance of pre-trained models,a re-ranking strategy based on graph convolutional networks is proposed.By leveraging graph convolution operators,this method propagates nearest neighbor features of pedestrians on the graph to refine the representation of each image,thereby enhancing retrieval outcomes.
person re-identificationpose estimation algorithmmulti-task lossgraph convolution oper-atorsre-ranking
万方数据
维普
