MonoSDF: Exploring Monocular Geometric Cues for Neural Implicit Surface Reconstruction

Yu et al. address a failure mode shared by NeuS (Wang et al., 2021), VolSDF (Yariv et al., 2021), and UNISURF (Oechsle et al., 2021): multi-view neural implicit surface methods produce excellent results on small, well-observed objects but degrade rapidly on large, textureless, or sparsely captured scenes. The diagnosis is that the photometric RGB loss alone is too weak a constraint in regions of low texture or low multi-view overlap.

MonoSDF augments the standard signed-distance-function plus volume- rendering pipeline with two monocular geometric cues, both produced by general-purpose off-the-shelf predictors that were never trained on the target scenes. A monocular depth predictor produces a per-pixel relative depth estimate, which is converted into a scale-and-shift-invariant depth-consistency loss against the rendered depth of the SDF. A monocular surface-normal predictor produces a per-pixel normal estimate, which is matched against the rendered surface normal (the analytic gradient of the SDF). Both losses are robust to the noise inherent in single-image predictions but provide dense supervision on every pixel.

The authors also conduct a systematic study of representation choice: monolithic MLPs versus single-resolution voxel grids versus the multi- resolution hash grids of Instant-NGP (Müller et al., 2022). The monocular cues improve every backbone, with the hash-grid variant producing the best speed-quality trade-off. On ScanNet, Replica, and Tanks-and-Temples the method substantially improves over NeuS and VolSDF on both small-scene and room-scale benchmarks.

MonoSDF crystallized the recipe — SDF backbone, volume rendering, hash-grid encoding, monocular priors — that subsequent indoor and outdoor surface-reconstruction work has largely converged on, including the Gaussian-Splatting surface variants that postdate this corpus.