An Adaptive Scaling Ensemble Matching Based Occluded Person Re-identification

Currently, the model performance is often not as effective as it should be due to the high similarity of different pedestrians due to occlusion and the easy misalignment of semantics across images. To overcome these obstacles, we propose an adaptive scaling ensemble matching based occluded person Re-identification. First, the overall ensemble matching thinking is adopted to generate the corresponding sets for the target image and the detection image through the designed ensemble generation module. Second, by generating a unique and distinguishable perceptual representation for each element in the target image collection, so that it is closer in proximity to positive exemplars while maintaining separation from negative instances. Finally, the significance of each channel is dynamically adjusted by the proposed channel adaptive scaling similarity in order to minimize the interference of too high or too low significance of the channel due to occlusion, and to effectively mitigate the semantic misalignment problem due to occlusion, so that the model still maintains a highly accurate matching in the face of occlusion environment. Extensive experiments conducted on the [Dataset Name] benchmark validate the proposed method’s superior performance against contemporary state-of-the-art techniques, while ablation studies confirm the individual contribution of each network component.