Research map / Point clouds / Point cloud classification
SPARS3R: Semantic Prior Alignment and Regularization for Sparse 3D Reconstruction
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.
SPARS3R aligns a dense MASt3R point cloud to a COLMAP sparse SfM point cloud via global RANSAC Procrustes alignment followed by SAM-guided semantic local alignment of outliers, then initializes Gaussian splatting for sparse-view novel view synthesis.
Problem
Sparse-view novel view synthesis with Gaussian splatting suffers from sparse initialization and inaccurate poses when dense depth-prior point clouds (DUSt3R/MASt3R) are used for initialization.
Approach
Fuse a dense prior point cloud with an accurate SfM point cloud via a two-stage alignment: a global RANSAC-based Procrustes alignment on triangulated correspondences, then piecewise rigid local alignment of outlier regions grouped into semantic masks produced by SAM prompted with outlier points.
Key results
- On 12-view Mip-NeRF 360, SPARS3R reaches 18.85 PSNR / 0.127 DSIM versus 16.23 PSNR for InstantSplat.
- On 12-view Tanks & Temples, SPARS3R achieves 29.90 PSNR and 0.007 DSIM, improving on InstantSplat's 26.97 PSNR.
- Ablation on Mip-NeRF 360 shows Global Fusion Alignment raises PSNR from 16.6 (SfM init) to 18.6, and Semantic Outlier Alignment adds a further 0.3 dB to 18.9.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SPARS3R instead of FSGS | Sparse-view NVS · Mip-NeRF 360 (12 views) | PSNR: 17.60 -> 18.85 | Helped |
| Use SPARS3R instead of InstantSplat | Sparse-view NVS · Tanks & Temples (12 views) | PSNR: 26.97 -> 29.90 | Helped |
| Use SPARS3R instead of FSGS | Sparse-view NVS · MVImgNet (12 views) | PSNR: 23.43 -> 25.85 | Helped |
| Use SPARS3R instead of InstantSplat | Sparse-view NVS · Tanks & Temples (3 views) | PSNR: 21.65 -> 21.83 | Helped |
| Use SPARS3R instead of FSGS | Sparse-view NVS · MVImgNet (3 views) | PSNR: 16.92 -> 18.62 | Helped |
| Use SPARS3R instead of InstantSplat | Sparse-view NVS · Tanks & Temples (6 views) | PSNR: 25.89 -> 26.80 | Helped |
| Use SPARS3R instead of InstantSplat | Sparse-view NVS · MVImgNet (6 views) | PSNR: 21.09 -> 22.75 | Helped |
| Initialize 3DGS with the MASt3R dense point cloud and poses instead of sparse SfM points and poses (ablation) | 3DGS · Mip-NeRF 360 | PSNR: 16.6 -> 15.9 | Hurt |
| Add Global Fusion Alignment to MASt3R-initialized 3DGS (ablation) | 3DGS · Mip-NeRF 360 | PSNR: 15.9 -> 18.6 | Helped |
| Add Semantic Outlier Alignment after Global Fusion Alignment (ablation) | 3DGS · Mip-NeRF 360 | PSNR: 18.6 -> 18.9 | Helped |
Surprising findings
- MASt3R-initialized 3DGS performs worse than SfM-initialized 3DGS on Mip-NeRF 360 (15.9 vs 16.6 PSNR), showing that denser but misaligned priors hurt rendering.
Limitations
- Relies on the quality of the semantic segmentation model; over- or under-segmentation degrades local alignment.
- Only piecewise rigid transforms are used; non-rigid alignment is left for future work.
- Evaluation uses test pose optimization for all methods, which adds time and may affect comparisons.
Experimental setup
Datasets: Tanks and Temples MVImgNet Mip-NeRF 360
Benchmarks: Mip-NeRF 360 Tanks & Temples MVImgNet
Metrics: PSNR SSIM LPIPS DreamSim RPE translation RPE rotation
Baselines: Instant-NGP 3DGS FSGS SparseGS DRGS CF-3DGS InstantSplat DUSt3R MASt3R
Builds on: 3D Gaussian Splatting MASt3R DUSt3R COLMAP Segment Anything
Closest papers by meaning
- SPAR3D: Stable Point-Aware Reconstruction of 3D Objects from Single Images (CVPR 2025 · similarity 0.90)
- Sparse-view Pose Estimation and Reconstruction via Analysis by Generative Synthesis (NeurIPS 2024 · similarity 0.89)
- PF3plat: Pose-Free Feed-Forward 3D Gaussian Splatting for Novel View Synthesis (ICML 2025 · similarity 0.89)
- Light3R-SfM: Towards Feed-forward Structure-from-Motion (CVPR 2025 · similarity 0.88)
- No Pose, No Problem: Surprisingly Simple 3D Gaussian Splats from Sparse Unposed Images (ICLR 2025 · similarity 0.87)
- Fast3R: Towards 3D Reconstruction of 1000+ Images in One Forward Pass (CVPR 2025 · similarity 0.87)
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.