Jointly adaptive cross-resolution person re-identification on super-resolution

Cross-resolution Person Re-identification (ReID) faces the significant challenge of large resolution variance across different camera views in real surveillance systems. Most approaches based on super-resolution (SR) excessively rely on the SR images, which may lead to the loss of low-resolution (LR) information. Meanwhile, the region-agnostic SR could pose interference to ReID. For this, we propose a jointly adaptive cross-resolution ReID framework that consists of a region-aware person super-resolution (RAPSR) and a resolution adaptive ReID (RAReID). RAPSR is equipped with spatial attention for enhancing crucial spatial regions in low-resolution (LR) images. RAReID extracts complementary features from LR and high-resolution (HR) images simultaneously and obtains more discriminative pedestrian representations through cascaded resolution adaptive feature fusion modules. Finally, by the joint training of RAPSR and RAReID, a greater ReID accuracy could be achieved. Extensive experiments demonstrate state-of-the-art performances on three derived and a native multi-resolution datasets.