Shape Topology-Driven Network for Unsupervised Keypoint Detection

3D keypoints provide an intuitive abstraction that facilitates the shape representation and plays a fundamental role in various downstream tasks. Although the existing methods tend to detect reliable and repeatable points, they only take the spatial relations between points into account and overlook the underlying topological structures of diverse shapes. Thus, we propose an unsupervised framework that leverages the skeletal outline of a shape to guide the detection of salient points. The skeletal representation enables the proposed network to capture the intrinsic topology of the point cloud. We first employ a DGCNN-like encoder to extract the point-wise features and then form a mask to predict the locations of keypoints. To take advantage of the neighboring information in skeletonization, we introduce a local topology construction strategy that associates the detected points with the regional structures. Finally, the skeletal outline is constructed by rationally connecting the keypoints and applied to refine the keypoint detection. Extensive experiments are conducted to demonstrate the effectiveness of our framework in detecting keypoints and producing expressive skeletal representations keypoints labels and are robust to noise.