Leaning compact and representative features for cross-modality person re-identification

This paper pays close attention to the cross-modality visible-infrared person re-identification (VI Re-ID) task, which aims match pedestrian samples between visible and infrared modes. In order reduce modality-discrepancy from different cameras, most existing works usually use constraints based on Euclidean metric. Because of distance metric strategy cannot effectively measure internal angles embedded vectors, solutions learn angularly discriminative feature embedding. Since important factor affecting classification task embedding vector is whether there an space, in this paper, we present a new loss function called Enumerate Angular Triplet (EAT) loss. Also, motivated by knowledge distillation, narrow down features modalities before embedding, further novel Cross-Modality Knowledge Distillation (CMKD) Benefit above two considerations, are enough way tackle problem. The experimental results RegDB SYSU-MM01 datasets have demonstrated that proposed method superior other advanced methods terms impressive performance. Code available at https://github.com/IVIPLab/LCCRF.