18 Mar 2024 | Sungphill Moon*, Hyeontae Son*, Dongcheol Hur, Sangwook Kim
GenFlow is an innovative approach for 6D pose estimation of novel objects, addressing the trade-off between accuracy and scalability. It leverages optical flow and 3D shape information to refine the pose iteratively, improving both generalization and accuracy. The method predicts optical flow between the rendered and observed images, refining the 6D pose through a PnP layer. By incorporating shape constraints and using an end-to-end differentiable system, GenFlow enhances performance. A cascade network architecture further improves multi-scale correlations and coarse-to-fine refinement. GenFlow achieved top rankings on unseen object pose estimation benchmarks in both RGB and RGB-D cases, outperforming existing state-of-the-art methods for seen objects without fine-tuning. The method's effectiveness is demonstrated through extensive experiments on various datasets, showing superior performance in handling novel objects.GenFlow is an innovative approach for 6D pose estimation of novel objects, addressing the trade-off between accuracy and scalability. It leverages optical flow and 3D shape information to refine the pose iteratively, improving both generalization and accuracy. The method predicts optical flow between the rendered and observed images, refining the 6D pose through a PnP layer. By incorporating shape constraints and using an end-to-end differentiable system, GenFlow enhances performance. A cascade network architecture further improves multi-scale correlations and coarse-to-fine refinement. GenFlow achieved top rankings on unseen object pose estimation benchmarks in both RGB and RGB-D cases, outperforming existing state-of-the-art methods for seen objects without fine-tuning. The method's effectiveness is demonstrated through extensive experiments on various datasets, showing superior performance in handling novel objects.