Hypergraph Convolutional Community based mostly Weakly Supervised Level Cloud Semantic Segmentation with Scene-Degree Annotations
Authors: Zhuheng Lu, Peng Zhang, Yuewei Dai, Weiqing Li, Zhiyong Su
Summary: Level cloud segmentation with scene-level annotations is a promising however difficult process. At present, the preferred approach is to make use of the category activation map (CAM) to find discriminative areas after which generate point-level pseudo labels from scene-level annotations. Nonetheless, these strategies all the time endure from the purpose imbalance amongst classes, in addition to the sparse and incomplete supervision from CAM. On this paper, we suggest a novel weighted hypergraph convolutional network-based technique, known as WHCN, to confront the challenges of studying point-wise labels from scene-level annotations. Firstly, in an effort to concurrently overcome the purpose imbalance amongst totally different classes and cut back the mannequin complexity, superpoints of a coaching level cloud are generated by exploiting the geometrically homogeneous partition. Then, a hypergraph is constructed based mostly on the high-confidence superpoint-level seeds that are transformed from scene-level annotations. Secondly, the WHCN takes the hypergraph as enter and learns to foretell high-precision point-level pseudo labels by label propagation. In addition to the spine community consisting of spectral hypergraph convolution blocks, a hyperedge consideration module is discovered to regulate the weights of hyperedges within the WHCN. Lastly, a segmentation community is educated by these pseudo level cloud labels. We comprehensively conduct experiments on the ScanNet and S3DIS segmentation datasets. Experimental outcomes display that the proposed WHCN is efficient to foretell the purpose labels with scene annotations, and yields state-of-the-art outcomes locally. The supply code is on the market at http://zhiyongsu.github.io/Project/WHCN.html.
