Object 6-DoF pose estimation using auxiliary learning
In order to accurately estimate the position and pose of an object in the camera coordinate sys-tem in challenging scenes with severe occlusion and scarce texture,while also enhancing network efficien-cy and simplifying the network architecture,this paper proposed a 6-DoF pose estimation method using auxiliary learning based on RGB-D data.The network took the target object image patch,corresponding depth map,and CAD model as inputs.First,a dual-branch point cloud registration network was used to obtain predicted point clouds in both the model space and the camera space.Then,for the auxiliary learn-ing network,the target object image patch and the Depth-XYZ obtained from the depth map were input to the multi-modal feature extraction and fusion module,followed by coarse-to-fine pose estimation.The es-timated results were used as priors for optimizing the loss calculation.Finally,during the performance evaluation stage,the auxiliary learning branch was discarded and only the outputs of the dual-branch point cloud registration network are used for 6-DoF pose estimation using point pair feature matching.Experi-mental results indicate that the proposed method achieves AUC of 95.9%and ADD-S<2 cm of 99.0%in the YCB-Video dataset;ADD(-S)result of 99.4%in the LineMOD dataset;and ADD(-S)result of 71.3%in the LM-O dataset.Compared with existing 6-DoF pose estimation methods,our method using auxiliary learning has advantages in terms of model performance and significantly improves pose estimation accuracy.
6-DoF pose estimationauxiliary learningRGB-D image3D point cloud
万方数据
维普
