SIANet: Support Information-Aware Network for Category-Agnostic Pose Estimation

Abstract

Category-Agnostic Pose Estimation (CAPE) aims to estimate the pose of arbitrary categories based on a few examples with keypoint definitions. Existing methods leverage a similarity matching process between support keypoint features and query image features to achieve CAPE. However, these methods mainly focus on extracting query image features, ignoring the rich prior information contained in support samples. In this paper, we propose a Support Information-Aware Network (SIANet) for CAPE. On the one hand, SIANet adopts keypoints re-parameterization technology to guide the extraction of visual features from the support image. On the other hand, SIANet performs support keypoint interaction to model category-specific prior structural information. Experiments show that our method outperforms previous state-of-the-art methods by a large margin.