3D Spatial Recognition Without Spatially Labeled 3D

Figure 1: (Left) Our framework, WyPR, jointly learns semantic segmentation and object detection for point cloud data from only scenelevel class tags. We find that encouraging consistency between the two tasks is key. (Right) Sample segmentation results from ScanNet val set, without seeing any point-level labels during training. Please refer to § 4.4 and Appendix F for more analysis and visualizations.