GIFS: Neural Implicit Function for General Shape Representation

TL;DR

The paper tackles the limitation of traditional neural implicit functions that rely on inside/outside partitions by introducing GIFS, a generalized implicit function that models pairwise point relations via a binary flag $b(\boldsymbol{\ p}_1, \boldsymbol{\ p}_2, \mathcal{S})$, enabling representation of non-watertight and multi-layer shapes. GIFS embeds points in a multi-scale voxel grid, fuses pair features with a max operation, and decodes to predict the pairwise flag, with an optional UDF branch to improve spatial cues; surfaces are extracted with an adapted Marching Cubes and subsequently refined by minimizing the UDF on the mesh. Experiments on ShapeNet show GIFS achieves state-of-the-art or competitive results for both watertight and general shapes, while offering substantial improvements in mesh quality and extraction efficiency over prior methods. This work broadens neural implicit modeling to general shapes and provides a practical path from GIFS to explicit meshes suitable for rendering and downstream tasks.

Abstract

Recent development of neural implicit function has shown tremendous success on high-quality 3D shape reconstruction. However, most works divide the space into inside and outside of the shape, which limits their representing power to single-layer and watertight shapes. This limitation leads to tedious data processing (converting non-watertight raw data to watertight) as well as the incapability of representing general object shapes in the real world. In this work, we propose a novel method to represent general shapes including non-watertight shapes and shapes with multi-layer surfaces. We introduce General Implicit Function for 3D Shape (GIFS), which models the relationships between every two points instead of the relationships between points and surfaces. Instead of dividing 3D space into predefined inside-outside regions, GIFS encodes whether two points are separated by any surface. Experiments on ShapeNet show that GIFS outperforms previous state-of-the-art methods in terms of reconstruction quality, rendering efficiency, and visual fidelity. Project page is available at https://jianglongye.com/gifs .

Figures (11)

• Figure 1: Comparison of different methods. Previous works divide the space into inside and outside like (a). One possible approach for multi-layer shape is to perform multi-class classification like (b). Our method no longer assigns predefined labels to points but classifies whether two points are separated by any surface, as shown in (c). In (d) and (e), we present a demo shape and a reconstructed bus from ShapeNet respectively.
• Figure 2: Generalized shape representation. In (a), we show examples of the binary flag. The flag of the points separated by the surface is positive, otherwise, it is negative. In (b) and (c), we show how a double-layer circle and a non-watertight curve are represented. The positive flags are bolded red.
• Figure 3: Overview of GIFS. We learn a neural implicit function to classify whether two points are separated by any surface. First, the input point cloud is encoded to a grid of features. Given two 3D points, corresponding features are extracted from the grid. Then a maxpool operator is utilized to keep the permutation invariance. The decoder takes the fused feature as input and approximates the binary flag between points. An extra UDF branch can be used to enhance the spatial perception of the feature.
• Figure 4: Comparison of methods trained on watertight shapes. Our method can reconstruct watertight shapes with the same accuracy as the state-of-the-art method.
• Figure 5: Reconstruction results of different methods on general shapes. Our method can reconstruct smooth, continuous surfaces, and achieve a better visual effect than previous methods.