RGB-Based Category-Level Object Pose Estimation via Depth Recovery and Adaptive Refinement

Category-level pose estimation methods have received widespread attention as they can be generalized to intra-class unseen objects. Although RGB-D-based category-level methods have made significant progress, reliance on depth image limits practical application. RGB-based methods offer a more practical and cost-effective solution. However, current RGB-based methods struggle with object geometry perception, leading to inaccurate pose estimation. We propose depth recovery and adaptive refinement for category-level object pose estimation from a single RGB image. We leverage DINOv2 to reconstruct the coarse scene-level depth from the input RGB image and propose an adaptive refinement network based on an encoder-decoder architecture to dynamically improve the predicted coarse depth and reduce its gap from the ground truth. We introduce a 2D-3D consistency loss to ensure correspondence between the point cloud obtained from depth projection and the objects in the 2D image. This consistency supervision enables the model to maintain alignment between the depth image and the point cloud. Finally, we extract features from the refined point cloud and feed them into two confidence-aware rotation regression branches and a translation and size prediction residual branch for end-to-end training. Decoupling the rotation matrix provides a more direct representation, which facilitates parameter optimization and gradient propagation. Extensive experiments on the REAL275 and CAMERA25 datasets demonstrate the superior performance of our method. Real-world estimation and robotic grasping experiments demonstrate our model robustness to occlusion, clutter environments, and low-textured objects.