BayesSDF is a novel probabilistic framework for uncertainty estimation in neural network-based implicit 3D representations. To overcome the limitations of existing neural network implicit surface models, which fail to provide a systematic method for quantifying uncertainty, it utilizes the Signed Distance Function (SDF) to provide a continuous and differentiable surface representation. We apply a Laplace approximation to the SDF weights and estimate local geometric instability using a Hessian-based metric. We experimentally demonstrate that the uncertainty estimates exhibit a strong correlation with surface reconstruction errors on synthetic and real-world data benchmarks. Consequently, BayesSDF lays the foundation for more robust, interpretable, and practical 3D perception systems.
