Topology-Aware Keypoint Detection via Skeleton-Based Shape Matching

3D keypoint detection endeavors to identify well-aligned and semantically consistent elements that reflect object shapes within point clouds, which plays a significant role in wide-ranging applications such as navigation and object tracking based on mobile devices. While existing approaches prioritize either salient features or statistic distributions for alignment, they overlook the underlying spatial topology of shapes. Although some recent methods take potential skeletons into account, they fail to associate this representation with local and global topology, thus reconciling comprehensive coverage and semantic awareness. To address this, we reckon keypoint detection as the skeleton-based shape matching and propose a two-branch framework that explicitly localizes the keypoints with broad coverage and semantic coherence in an unsupervised manner. Specifically, one branch incorporates the keypoint detector with a skeleton generator to infer the coarse skeletons that represent the global topology. Meanwhile, another branch leverages skeletal sphere estimation to generate the skeletal point set that sustains the local structures, serving as the foundation for optimizing the skeletons formed by keypoints. Since these skeletal representations capture both the structural essence and semantic attributes of a shape, our model is capable of extracting semantically rich keypoints with good alignment. We extensively evaluate our method on different datasets to demonstrate its effectiveness and competitiveness in 3D keypoint detection.