An Appearance-Based Keypoint Selection Method for 6-DoF Pose Estimation of Irregular-Shaped Objects

This paper proposes an appearance based method for keypoint selection, with the goal of improving pose estimation and supporting downstream tasks. In conventional pose estimation pipelines, keypoints are typically detected first, and the pose of the target object is subsequently computed using specialized algorithms. Traditional keypoint selection methods, such as those based on bounding boxes, rely on a fixed spatial arrangement of keypoints. However, for objects with non-standard geometries, pose estimation accuracy can be significantly enhanced when keypoints are adaptively selected from the object's surface. This paper introduces a new approach for keypoint selection based on self-supervised learning, incorporating multi-modal information to enrich feature extraction. By integrating depth data with RGB images, the proposed method better captures the geometric characteristics of objects. Unlike traditional methods, our approach dynamically selects keypoints based on the specific shape of the object. This method aims to provide more accurate inputs for pose estimation tasks, thereby improving the overall performance of pose estimation, particularly for objects with complex shapes.