DeepSDF: Learning Continuous Signed Distance Functions for Shape\n Representation

Abstract

Computer graphics, 3D computer vision and robotics communities have produced\nmultiple approaches to representing 3D geometry for rendering and\nreconstruction. These provide trade-offs across fidelity, efficiency and\ncompression capabilities. In this work, we introduce DeepSDF, a learned\ncontinuous Signed Distance Function (SDF) representation of a class of shapes\nthat enables high quality shape representation, interpolation and completion\nfrom partial and noisy 3D input data. DeepSDF, like its classical counterpart,\nrepresents a shape's surface by a continuous volumetric field: the magnitude of\na point in the field represents the distance to the surface boundary and the\nsign indicates whether the region is inside (-) or outside (+) of the shape,\nhence our representation implicitly encodes a shape's boundary as the\nzero-level-set of the learned function while explicitly representing the\nclassification of space as being part of the shapes interior or not. While\nclassical SDF's both in analytical or discretized voxel form typically\nrepresent the surface of a single shape, DeepSDF can represent an entire class\nof shapes. Furthermore, we show state-of-the-art performance for learned 3D\nshape representation and completion while reducing the model size by an order\nof magnitude compared with previous work.\n