Hypergraph Convolutional Neighborhood based Weakly Supervised Degree Cloud Semantic Segmentation with Scene-Diploma Annotations
Authors: Zhuheng Lu, Peng Zhang, Yuewei Dai, Weiqing Li, Zhiyong Su
Abstract: Degree cloud segmentation with scene-level annotations is a promising nonetheless troublesome course of. At current, the popular strategy is to utilize the class activation map (CAM) to search out discriminative areas after which generate point-level pseudo labels from scene-level annotations. Nonetheless, these methods on a regular basis endure from the aim imbalance amongst lessons, along with the sparse and incomplete supervision from CAM. On this paper, we propose a novel weighted hypergraph convolutional network-based approach, often known as WHCN, to confront the challenges of learning point-wise labels from scene-level annotations. Firstly, in an effort to concurrently overcome the aim imbalance amongst completely totally different lessons and in the reduction of the model complexity, superpoints of a training stage cloud are generated by exploiting the geometrically homogeneous partition. Then, a hypergraph is constructed based mostly totally on the high-confidence superpoint-level seeds which are reworked from scene-level annotations. Secondly, the WHCN takes the hypergraph as enter and learns to predict high-precision point-level pseudo labels by label propagation. Along with the backbone group consisting of spectral hypergraph convolution blocks, a hyperedge consideration module is found to manage the weights of hyperedges inside the WHCN. Lastly, a segmentation group is educated by these pseudo stage cloud labels. We comprehensively conduct experiments on the ScanNet and S3DIS segmentation datasets. Experimental outcomes show that the proposed WHCN is environment friendly to predict the aim labels with scene annotations, and yields state-of-the-art outcomes domestically. The provision code is available on the market at http://zhiyongsu.github.io/Project/WHCN.html.
