Sparse keypoint models for 6D object pose estimation

In this paper, we present an approach to generate sparse object models for keypoint-based 6D object pose estimation. Keypoint-based object models usually consist of thousands of keypoints. Our approach generates sparse models by identifying and removing keypoints that are not relevant to the object localization. It applies data association to detect duplicate keypoints and applies statistical analysis to identify keypoints that have not been detected reliably during model generation. Our approach furthermore ensures that keypoints are well distributed across the volume of the object model. We evaluated our approach using a SIFT-based 6D object localization system on the basis of real world datasets. In our experiments, we achieved a reduction of the model sizes to approximately 1% of the original model size without a substantial loss of localization performance.