DeepSDF: Learning Continuous Signed Distance Functions for Shape Representation

DeepSDF introduces a learned, continuous Signed Distance Function representation for 3D shapes, aimed at high-quality shape representation, interpolation, and completion from partial and noisy 3D input. As in the classical SDF, a shape's surface is represented by a continuous volumetric field: the magnitude at a point gives the distance to the surface boundary and the sign indicates whether the point is inside or outside the shape, so the surface is implicitly encoded as the zero-level-set of the learned function while space is explicitly classified as interior or exterior.

Unlike classical analytical or discretized voxel SDFs, which typically represent the surface of a single shape, DeepSDF learns to represent an entire class of shapes with one model. The authors report state-of-the-art performance for learned 3D shape representation and completion while reducing model size by an order of magnitude compared with previous work. This combination of continuous representation, generalization across a shape class, and compactness made it an influential approach to implicit neural shape representation.