Pose Knowledge distill guidance: Effective Pose guide learning for Person Re-Identification

The robust human representation embedding is the key challenge of the person-identification task. Although the previous methods achieve competitive results, they still depend on the appearance feature which can be failed in the pedestrian surveillance camera due to the changing of the human's appearance. To overcome this limitation, we introduce Pose Knowledge distill guidance, an effective pose guide learning for the Person Re-Identification framework that leverages pose knowledge guidance through Knowledge distillation to yield more generalized feature embeddings. Because of the significance of human pose in distinguishing individuals, apart from the appearance knowledge, our framework also guides the model to learn the pose information, consequently reducing the overreliance on the appearance of the human and thereby enhancing the model's robustness. Furthermore, the pose knowledge guides transfer the pose knowledge to the Re-Identification model without no increasing the computation cost for the backbone. Through our framework, the re-identification model demonstrates competitive performance, as evidenced by its state-of-the-art results on two datasets: Market-1501 and CuHK03. These results attest to the effectiveness of our approach.