Reproduce “Gen6D: Generalizable Model-Free 6-DoF Object Pose Estimation from RGB Images” on OpenResearch

About this paper

Yuan Liu, Yilin Wen, Sida Peng, Cheng Lin, Xiaoxiao Long, Taku Komura, Wenping Wang

Gen6D proposes a model-free 6-DoF object pose estimation method that relies solely on RGB images and a few reference views to estimate poses for arbitrary, unseen objects. The system achieves an average ADD-0.1d of 50.39% and Prj-5 of 82.64% on the new GenMOP dataset for novel objects, outperforming existing generalizable and some instance-specific methods.

Key insights

• Contextual normalization across all reference images and a Transformer architecture significantly enhance the robustness and accuracy of viewpoint selection, even with sparse reference views and cluttered backgrounds.
• A 3D feature volume-based refinement approach enables model-free pose optimization by directly reasoning about 3D pose changes within a bounded feature space, eliminating the need for 3D models or rendering.
• Normalizing the object coordinate system and warping reference images to a canonical view allows for robust feature learning and improved generalization across diverse objects.

Summary by alphaXiv.