Omni-supervised Point Cloud Segmentation via Gradual Receptive Field Component Reasoning

Hidden features in neural network usually fail to learn informative\nrepresentation for 3D segmentation as supervisions are only given on output\nprediction, while this can be solved by omni-scale supervision on intermediate\nlayers. In this paper, we bring the first omni-scale supervision method to\npoint cloud segmentation via the proposed gradual Receptive Field Component\nReasoning (RFCR), where target Receptive Field Component Codes (RFCCs) are\ndesigned to record categories within receptive fields for hidden units in the\nencoder. Then, target RFCCs will supervise the decoder to gradually infer the\nRFCCs in a coarse-to-fine categories reasoning manner, and finally obtain the\nsemantic labels. Because many hidden features are inactive with tiny magnitude\nand make minor contributions to RFCC prediction, we propose a Feature\nDensification with a centrifugal potential to obtain more unambiguous features,\nand it is in effect equivalent to entropy regularization over features. More\nactive features can further unleash the potential of our omni-supervision\nmethod. We embed our method into four prevailing backbones and test on three\nchallenging benchmarks. Our method can significantly improve the backbones in\nall three datasets. Specifically, our method brings new state-of-the-art\nperformances for S3DIS as well as Semantic3D and ranks the 1st in the ScanNet\nbenchmark among all the point-based methods. Code will be publicly available at\nhttps://github.com/azuki-miho/RFCR.\n