Identity Hides in Darkness: Learning Feature Discovery Transformer for Nighttime Person Re-Identification

Person re-identification (Re-ID) aims to retrieve all images of the specific person captured by non-overlapping cameras and scenarios. Regardless of the significant success achieved by daytime person Re-ID methods, they will perform poorly due to the degraded imaging quality under low-light conditions. Therefore, some works attempt to synthesize low-light images to explore the challenges in the nighttime, which omits the fact that synthetic images may not realistically reflect the challenges of person Re-ID at night. Moreover, other works follow the "enhancement-then-match" manner, but it is still hard to capture discriminative identity features owing to learning enlarged irrelevant noise for identifying pedestrians. To this end, we propose a novel nighttime person Re-ID method, termed Feature Discovery Transformer (FDT), explicitly capturing the pedestrian identity information hidden in darkness at night. More specifically, the proposed FDT model contains two novel modules: the Frequency-wise Reconstruction Module (FRM) and the Attribute Guide Module (AGM). In particular, to reduce noise disturbance and discover pedestrian identity details, the FRM utilizes the Discrete Haar Wavelet Transform to acquire the high- and low-frequency components for learning person features. Furthermore, to avoid high-frequency components being over-smoothed by low-frequency ones, we propose a novel Normalized Contrastive Loss (NCL) to help the model obtain the identity details in high-frequency components for extracting discriminative person features. Then, to further decrease the negative bias caused by appearance-irrelevant features and enhance the pedestrian identity features, the AGM improves the robustness of the learned features by integrating the auxiliary information, i.e., camera ID and viewpoint. Extensive experimental results demonstrate that our proposed FDT model can achieve state-of-the-art performance on two realistic nighttime person Re-ID benchmarks, i.e., Night600 and RGBNT201rgb datasets.