DeepSDF: Learning Continuous Signed Distance Functions for Shape Representation

TL;DR

DeepSDF introduces a continuous, latent-conditioned SDF representation for 3D shapes, enabling high-fidelity surfaces, interpolation, and completion from partial data with a compact memory footprint. It uses an encoder-less auto-decoder framework to learn a latent space of shapes, allowing a single network to represent thousands of shapes via per-shape latent codes and a shared decoder. The approach achieves state-of-the-art results on shape reconstruction and completion tasks and supports smooth latent-space interpolation, demonstrating strong generalization and efficient implicit geometry. Practical limitations include inference time due to optimization and the need for a canonical pose, guiding future work toward faster optimization and broader scene modeling.

Abstract

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

Figures (21)

• Figure 1: DeepSDF represents signed distance functions (SDFs) of shapes via latent code-conditioned feed-forward decoder networks. Above images are raycast renderings of DeepSDF interpolating between two shapes in the learned shape latent space. Best viewed digitally.
• Figure 2: Our DeepSDF representation applied to the Stanford Bunny: (a) depiction of the underlying implicit surface $SDF=0$ trained on sampled points inside $SDF<0$ and outside $SDF>0$ the surface, (b) 2D cross-section of the signed distance field, (c) rendered 3D surface recovered from $SDF=0$. Note that (b) and (c) are recovered via DeepSDF.
• Figure 3: In the single-shape DeepSDF instantiation, the shape information is contained in the network itself whereas the coded-shape DeepSDF, the shape information is contained in a code vector that is concatenated with the 3D sample location. In both cases, DeepSDF produces the SDF value at the 3D query location,
• Figure 4: Different from an auto-encoder whose latent code is produced by the encoder, an auto-decoder directly accepts a latent vector as an input. A randomly initialized latent vector is assigned to each data point in the beginning of training, and the latent vectors are optimized along with the decoder weights through standard backpropagation. During inference, decoder weights are fixed, and an optimal latent vector is estimated.
• Figure 5: Compared to car shapes memorized using OGN ogn2017 (right), our models (left) preserve details and render visually pleasing results as DeepSDF provides oriented surace normals.