Cross-modality video person Re-ID with modality-aware cosine-triplet loss

Abstract

Person Re-Identification (Re-ID) is one of the important applications for surveillance. The scenarios where we need to identify the subjects captured by night-vision (infrared) cameras are a significant challenge to the existing Re-ID techniques, where only color footage is available for comparison. This is due to large differences in the composition between color and infrared images, which results in appearance-based information becoming less reliable for Re-ID. For this reason, we hypothesized that motion information from sequences of inputs is vital for cross-modality (visible-to-infrared) Re-ID. From our initial findings, motion information from the sequence of frames significantly improved the cross-modality Re-ID performance. In addition, choices of distance metrics (Euclidean vs. cosine) have a significant effect on the overall performance. As a result, the experimental performance on SYSU-MM01 reached 72.70% in mAP and 73.27% in rank-1 accuracy and yielded significant performance gains of 28.32% in mAP and 29.14% in rank-1 accuracy over our baseline. The performance competes with the existing state-of-the-art techniques tested on the same dataset.