2019/10/02 by G. K. Sharma, Sharma, Gopal, Evangelos Kalogerakis +3
Computer Science · Earth and Planetary Sciences · Engineering · #3D Shape Modeling and Analysis #3D Surveying and Cultural Heritage #Computer Graphics and Visualization Techniques #Computer Vision and Pattern Recognition (cs.CV) #FOS: Computer and information sciences #Machine Learning (cs.LG)
paper · pdf · doi:10.48550/arxiv.1910.01269
openalex publication_date 2019/10/02 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28
User generated 3D shapes in online repositories contain rich information\nabout surfaces, primitives, and their geometric relations, often arranged in a\nhierarchy. We present a framework for learning representations of 3D shapes\nthat reflect the information present in this meta data and show that it leads\nto improved generalization for semantic segmentation tasks. Our approach is a\npoint embedding network that generates a vectorial representation of the 3D\npoints such that it reflects the grouping hierarchy and tag data. The main\nchallenge is that the data is noisy and highly variable. To this end, we\npresent a tree-aware metric-learning approach and demonstrate that such learned\nembeddings offer excellent transfer to semantic segmentation tasks, especially\nwhen training data is limited. Our approach reduces the relative error by\n10.2 % with 8 training examples, by 11.72 % with 120 training examples\non the ShapeNet semantic segmentation benchmark, in comparison to the network\ntrained from scratch. By utilizing tag data the relative error is reduced by\n12.8 % with 8 training examples, in comparison to the network trained from\nscratch. These improvements come at no additional labeling cost as the meta\ndata is freely available.\n