Research map / 3D vision / Depth and stereo
Sparis: Neural Implicit Surface Reconstruction of Indoor Scenes from Sparse Views
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Sparis reconstructs indoor scene surfaces from sparse views by adding triangulated inter-image depth priors from feature matching, a cross-view reprojection loss, and angular and epipolar filters on top of a VolSDF/NeuRIS-style SDF pipeline.
Problem
Indoor surface reconstruction from sparse views (15-20 images on ScanNet, 10 on Replica) where monocular depth priors suffer scale ambiguity and the geometry collapses.
Approach
Replace scale-ambiguous monocular depth supervision with absolute depth obtained by triangulating matched pixel pairs from a feature matching network (RoMa), enforce cross-view consistency with a reprojection loss, and filter noisy matches using an angular view-selection score and a Sampson-distance-based epipolar weight.
Key results
- On 10 ScanNet scenes with 15-20 views, Sparis reaches F-score 0.647 versus 0.464 for the next best method (NeuRIS).
- On 8 Replica scenes with 10 views, Sparis reaches F-score 0.825 versus 0.454 for MonoSDF.
- Ablation on ScanNet: adding the inter-image depth loss and reprojection loss to the normal prior raises F-score from 0.253 to 0.647.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Sparis instead of NeuRIS | Indoor surface reconstruction · ScanNet, 15–20 input views | F-score: 0.464 -> 0.647 | Helped |
| Use Sparis instead of MonoSDF | Indoor surface reconstruction · Replica, 10 input views | F-score: 0.454 -> 0.825 | Helped |
| Add normal priors to neural rendering without priors (ablation) | Sparis · ScanNet, 10 scenes | F-score: 0.244 -> 0.253 | Helped |
| Add inter-image depth loss (ablation) | Sparis with normal priors · ScanNet, 10 scenes | F-score: 0.253 -> 0.598 | Helped |
| Add cross-view reprojection loss (ablation) | Sparis with normal priors and inter-image depth loss · ScanNet, 10 scenes | F-score: 0.598 -> 0.647 | Helped |
| Add inter-image depth loss (ablation) | Sparis with normal priors · ScanNet, 10 scenes | F-score: 0.253 -> 0.560 | Helped |
| Add cross-view reprojection loss (ablation) | Sparis with normal priors · ScanNet, 10 scenes | F-score: 0.253 -> 0.560 | Helped |
| Add epipolar weight (ablation) | Sparis with angular filter · ScanNet, 10 scenes | F-score: 0.617 -> 0.647 | Helped |
| Add angular filter (ablation) | Sparis with epipolar weight · ScanNet, 10 scenes | F-score: 0.624 -> 0.647 | Helped |
Surprising findings
- Monocular depth supervision alone in the sparse setting degrades reconstruction, motivating the matching-based depth prior.
Limitations
- Reconstruction quality depends on the accuracy of the feature matching network, so matching errors propagate into the priors.
- MonoSDF and NeuS needed hyperparameter changes or failed to produce meshes for some scenes, so baseline comparisons required tuning and partial scene reporting.
Experimental setup
Datasets: ScanNet Replica
Benchmarks: ScanNet Replica
Metrics: F-score Accuracy Completeness Precision Recall
Baselines: COLMAP TransMVSNet DDP-NeRF DäRF NeuS VolRecon HelixSurf S^3P MonoSDF NeuRIS
Builds on: VolSDF NeuRIS MonoSDF RoMa
Closest papers by meaning
- SparseGNV: Generating Novel Views of Indoor Scenes with Sparse RGB-D Images (AAAI 2024 · similarity 0.89)
- NeuSurf: On-Surface Priors for Neural Surface Reconstruction from Sparse Input Views (AAAI 2024 · similarity 0.87)
- NeRFPrior: Learning Neural Radiance Field as a Prior for Indoor Scene Reconstruction (CVPR 2025 · similarity 0.87)
- NC-SDF: Enhancing Indoor Scene Reconstruction Using Neural SDFs with View-Dependent Normal Compensation (CVPR 2024 · similarity 0.86)
- MP-SfM: Monocular Surface Priors for Robust Structure-from-Motion (CVPR 2025 · similarity 0.86)
- Sparse3D: Distilling Multiview-Consistent Diffusion for Object Reconstruction from Sparse Views (AAAI 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.