Abstract
NeuSurf is a novel framework for high-quality 3D surface reconstruction from sparse input views (as few as three images). By leveraging "on-surface priors"—geometric data extracted directly from input views via Structure from Motion (SfM)—it eliminates the need for expensive pre-training on massive datasets while outperforming state-of-the-art methods in sparse-view scenarios.
Introduction
Neural implicit functions like NeuS or NeRF represent surfaces as continuous fields but typically require dense image sets. With sparse inputs (e.g., 3 views), these models suffer from shape-radiance ambiguity, where the model creates noisy, physically nonsensical geometry that only looks correct from training angles. Existing solutions often rely on generalizable priors pre-trained on large datasets, which can be brittle and fail under unusual camera configurations. NeuSurf provides a robust, training-free alternative by deriving priors directly from the scene being reconstructed.
Methodology
NeuSurf introduces a two-pronged approach to guide neural reconstruction using points present in the input data via SfM.
The framework consists of three main components:
- On-Surface Global Geometry Alignment: Instead of direct depth supervision from sparse SfM points, NeuSurf trains an Unsigned Distance Function (UDF) network ($f_\theta$). UDFs are more flexible than Signed Distance Functions (SDFs) for representing the partial, open surfaces typical of SfM outputs. This UDF acts as a continuous "coarse map" to stabilize early optimization. To train the UDF without ground truth distance values, the authors use a "moving operation" to pull query points toward the surface.
- On-Surface Local Geometry Refinement: To capture fine details, the model enforces feature consistency by projecting surface points onto both seen and unseen views, ensuring matching features across perspectives.
- Pseudo On-Surface Points: To address uneven SfM point distribution, the model identifies "pseudo-points" at ray-surface intersections, densifying supervision during optimization.
Experiments and Results
NeuSurf was evaluated on the DTU and BlendedMVS datasets, focusing on sparse settings with only 3 views per scene.
NeuSurf significantly outperformed baselines like SparseNeuS, VolRecon, and MonoSDF. On the DTU "Little-overlap" setting, NeuSurf achieved a mean Chamfer Distance (CD) of 1.35, compared to 3.02 for VolRecon and 1.86 for MonoSDF. Qualitatively, NeuSurf maintains structural integrity and fine details (e.g., sharp building edges) while other methods produce "broken" or "melted" surfaces. Ablation studies confirmed that both the global UDF prior and local feature consistency are essential for high-quality sparse reconstruction.
Conclusion and Limitations
While NeuSurf achieves superior quality, it requires per-scene optimization (approx. 300k iterations on an NVIDIA RTX 3090) and depends on the initial SfM point cloud quality. If input images are featureless, the on-surface prior will be weak. Future work may focus on accelerating optimization and improving performance in low-texture regions.
Summary updated