Part-Pose Guided Amur Tiger Re-Identification

Cen Liu, Rong Zhang, Lijun Guo; Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019, pp. 0-0

In this paper, we present our solution to tiger re-identification (re-ID) in both the plain and the wild tracks in the 2019 Computer Vision for Wild life Conservation Challenge (CVWC2019). We introduce a novel part-pose guided framework for the tiger re-ID task, which consists of two part streams and one full stream based on the pose characteristics of tiger. Considering missing and inaccurate pose annotations, the two part streams are used as a regulator to guide the full stream in learning and aligning the local features in the training stage. We only use the learnt full stream for the tiger re-ID task in the inference stage. The proposed model has the advantage that despite requiring pose information at training time it is not needed during inference, so it is particularly suitable for tiger re-ID in the wild. Our proposed method outperforms the state-of-the-art and finished top in both the PlainID and WildID competitions at CVWC2019. The source of code will be public available at https://github.com/LcenArthas/CVWC2019-Amur-Tiger-Re-ID