A double-covered probabilistic ray-based neural shape representation

Solutions to visual computing problems such as shape reconstruction, analysis, generation, and retrieval rely on fast and task-appropriate 3D shape representations for efficiency. This work explores representing shapes with neural ray fields, which are inherently more efficient to render than 3D coordinate-based neural representations requiring only a single network evaluation per ray, but with high data requirements and reduced fidelity thus far. We propose an improved ray field representation with improved gradient flow and a more tractable parameter domain, along with a more stable objective function and data augmentation improving stability and fidelity. We study the strengths and failure modes of the prior art and our method, and how performance scales with data and compute.