Pose for Everything: Towards Category-Agnostic Pose Estimation

Existing works on 2D pose estimation mainly focus a certain category, e.g. human, animal, and vehicle. However, there are lots of application scenarios that require detecting the poses/keypoints unseen class objects. In this paper, we introduce task Category-Agnostic Pose Estimation (CAPE), which aims to create model capable any object given only few samples with keypoint definition. To achieve goal, formulate problem as matching design novel CAPE framework, termed POse Matching Network (POMNet). A transformer-based Keypoint Interaction Module (KIM) is proposed capture both interactions among different keypoints relationship between support query images. We also Multi-category (MP-100) dataset, dataset 100 categories containing over 20K instances well-designed for developing algorithms. Experiments show our method outperforms other baseline approaches by large margin. Codes data available at https://github.com/luminxu/Pose-for-Everything .