The paper presents a novel approach to category-agnostic pose estimation, aiming to enable keypoint localization for any object category using a single model. This method leverages geometrical relations between keypoints through a Graph Transformer Decoder, enhancing the accuracy of keypoint localization. The method was validated on the MP-100 benchmark, outperforming previous techniques significantly. The approach demonstrates scalability and efficiency, marking a substantial departure from conventional techniques that treat keypoints as isolated entities.
Publication date: 29 Nov 2023
Project Page: https://orhir.github.io/pose-anything/
Paper: https://arxiv.org/pdf/2311.17891
