Pose-Guided Complementary Features Learning for Amur Tiger Re-Identification

Re-identifying different animal individuals is of significant importance to animal behavior and ecology research and protecting endangered species. This paper focuses on Amur tiger re-identification (re-ID) using computer vision (CV) technology. State-of-the-art CV-based Amur tiger re-ID methods extract local features from different body parts of tigers based on stand-alone pose estimation methods. Consequently, they are limited by the pose estimation accuracy and suffer from self-occluded body parts. Instead of estimating elaborated body poses, this paper simplifies tiger poses as right-headed or left-headed and utilizes this information as an auxiliary pose classification task to supervise the feature learning. To further enhance the feature discriminativeness, this paper learns multiple complementary features by steering different feature extraction network branches towards different regions of the tiger body via erasing activated regions from input tiger images. By fusing the pose-guided complementary features, this paper effectively improves the Amur tiger re-ID accuracy as demonstrated in the evaluation experiments on two test datasets. The code and data of this paper are publicly available at https://github.com/liuning-scu-cn/AmurTigerReID.