Improving 6D Object Pose Estimation Based on Semantic Segmentation

The performance of 6D pose estimation, which is important for scene understanding, can be improved by more accurate object segmentation. RGB-D data including depth maps can provide more accurate position information than RGB data for semantic segmentation. In this work, we propose a novel two-stage RGBD-based pose estimation network, which can provide more precise semantic segmentation and effective point cloud features. Firstly, we use a lightweight semantic segmentation head to process the RBG-D data to get the pixel-level clustered mask, and then use a multi-scale and attention-based backbone to extract the point cloud features for pose estimation. We analyze the performance of our network on the YCB-Video dataset and the results show that our method is comparable to current state-of-the-art methods after optimization.