Keypoints Estimation on Partial Point Clouds

Estimating 3D keypoints of an object has achieved considerable interest, especially in computer graphics and ma-chine vision. The complexity of the problem increases when the object is noisy, decimated, or partially occluded. To this end, we present an approach to estimate 3D keypoints from the partial PCDs such that they also highlight the object's missing parts. It is based on two modules: shape completion and keypoints estimation. The first module computes the object's missing parts and generates a dense shape which is used by the second module to estimate semantically and geometrically consistent keypoints. The approach also estimates confidence scores that allow the selection of valid keypoint considering the object's structure. We evaluate the proposed method for four different backbones and observe that we get better results using PointNet. Furthermore, the presented results show that the Percentage of Correct Keypoints (PCK) and Dual Alignment Score (DAS) for the keypoints estimated from the dense shapes (36.61 and 76.29) are 23.51 % and 12.14% higher, respectively, compared to those estimated from the partial objects (13.10 and 64.15).