We propose DTM, a novel difference-based temporal module to leverage historical information in category-level 6DoF pose tracking tasks. It can be easily integrated with various category-level 6DoF pose tracking models which use RGBD data as input. This module extracts temporal features through a KNN-based difference calculation strategy from both, pixels and 3D points. We evaluate this module on two pose estimation datasets, NOCS-REAL275 and MoVi-E by integrating our module with two state-of-the-art 6D pose tracking models. The result shows that DTM can significantly increase the accuracy and robustness of category-level 6DoF trackers.
Creative Commons Attribution 4.0 International License (CC-BY 4.0)