Neural Surface Localization for Unsigned Distance Fields

By providing a continuous and differentiable representation of three-dimensional shapes whose topology can change, Signed Distance Fields (SDFs) have proved their worth. However, they are best suited to modeling watertight surfaces. Non-watertight ones, such as those garments are made of, can be handled by meshing a non-zero level set, which amounts to wrapping an SDF around them, but at the cost of an accuracy loss.

Thus, Unsigned Distance Fields (UDFs) have emerged as an effective alternative for representing open surfaces. Unfortunately, when it becomes necessary to convert an implicit surface into an explicit one, for example for rendering purposes, UDFs are at a disadvantage. While there are well-established approaches such as Marching Cubes and Dual Contouring for triangulating SDF-based implicit surfaces, existing triangulation methods that can operate on UDFs are less reliable.

In an upcoming ECCV paper , we introduce a deep-learning based approach to surface localization in unsigned distance fields that avoids these pitfalls. We provide some results in this video.

We formulate the surface localization problem as a local cell-wise classification one. More specifically, we start with an UDF whose values and gradients are given on a 3D grid and train a network to turn them into an SDF-like field in which signs flip across boundaries and that can be triangulated using the standard Marching Cubes algorithm. To train the network, we use watertight surfaces for which an SDF ground-truth can be computed precisely. However, once trained, the network can run on non-watertight surfaces just as well because it makes local decisions.

Our approach is entirely data-driven, without any hand-crafted rules that can result in harmful biases. Thus, our contribution is a novel data-driven approach for neural open surface localization. It comes without any annotation cost. The network is easy to train, adapt, and scalable to various scenarios. We demonstrate on the ABC, ShapeNet-Cars, and MGN garments datasets that our approach outperforms the state-of-the-art MeshUDF and DualMesh-UDF algorithms, when used in conjunction with either Marching Cubes or Dual Contouring, respectively.