FS-Net: Fast Shape-based Network for Category-Level 6D Object Pose Estimation with Decoupled Rotation Mechanism

In this paper, we focus on category-level 6D pose and size estimation frommonocular RGB-D image. Previous methods suffer from inefficient category-levelpose feature extraction which leads to low accuracy and inference speed. Totackle this problem, we propose a fast shape-based network (FS-Net) withefficient category-level feature extraction for 6D pose estimation. First, wedesign an orientation aware autoencoder with 3D graph convolution for latentfeature extraction. The learned latent feature is insensitive to point shiftand object size thanks to the shift and scale-invariance properties of the 3Dgraph convolution. Then, to efficiently decode category-level rotationinformation from the latent feature, we propose a novel decoupled rotationmechanism that employs two decoders to complementarily access the rotationinformation. Meanwhile, we estimate translation and size by two residuals,which are the difference between the mean of object points and ground truthtranslation, and the difference between the mean size of the category andground truth size, respectively. Finally, to increase the generalizationability of FS-Net, we propose an online box-cage based 3D deformation mechanismto augment the training data. Extensive experiments on two benchmark datasetsshow that the proposed method achieves state-of-the-art performance in bothcategory- and instance-level 6D object pose estimation. Especially incategory-level pose estimation, without extra synthetic data, our methodoutperforms existing methods by 6.3% on the NOCS-REAL dataset.