<p>Category-level 6DoF pose estimation aims to predict the rotation, translation and scale of unseen objects for a given category. Most of the previous prior-free methods use RGB-D images as input, extract the images features and the point cloud features, and then estimate 6D poses by performing transformations using concatenated image and point cloud features. Given that the unstructured and disordered 3D point cloud makes the extraction of point cloud features not very efficient, we propose a new 6D pose estimation network, named PSG-6D, which aims to optimize point cloud learning by performing implicit space transformation from camera-space features to world-space features and establishing their correspondences without the need of a category-level prior. Specifically, we develop an SO(3)-equivariant network and a global enhancement block for encoding point cloud recovered from depth maps to extract SO(3)-equivariant features, enhancing the robustness of the features to rotation, and fusing the global features of the point cloud features respectively. At the same time, a multilayer perceptron is used to encode the positional information of the point cloud. Subsequently, the implicit space transformation from camera coordinates to world coordinates is performed by concatenating the image features and the geometric features containing the position information. Our method achieves state-of-the-art performance on both the CAMERA25 synthetic dataset and the challenging REAL275 real scene dataset. Our method also achieves satisfactory results on the challenging occlusion dataset HouseCat6D.</p>