Research map / 3D vision / Face and shape
AniGrad: Anisotropic Gradient-Adaptive Sampling for 3D Reconstruction From Monocular Video
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.
Proposes AniGrad, which re-formulates a 3D CNN TSDF predictor as a linear combination of learned local basis functions so per-cube gradient bounds give anisotropic adaptive sub-voxel sampling for faster mesh extraction from monocular video.
Problem
Image-based TSDF reconstruction with 3D CNNs needs high-resolution query grids for detailed surfaces, making surface extraction slow and meshes large.
Approach
Interpret each voxel's feature vector as weights over learned coordinate-MLP basis functions, precompute per-basis gradient extrema offline, and use these to analytically bound the local TSDF gradient and set anisotropic per-cube sample rates before querying the TSDF.
Key results
- Surface extraction latency of 1.11 s vs 24.81 s for FineRecon on ScanNet (about 22x reduction).
- Mesh size 5.47 MB vs 54.85 MB for FineRecon, with F-score 74.0 vs 74.1.
- Chamfer distance 5.29 cm vs FineRecon's 4.39 cm, i.e. second-best accuracy.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use adaptive anisotropic sub-sampling instead of FineRecon's uniform 1 cm sampling | AniGrad · ScanNet test set | F-score: 74.1 -> 74.0 | No clear effect |
| Use adaptive anisotropic sub-sampling instead of FineRecon's uniform 1 cm sampling | AniGrad · ScanNet test set | Mesh size (MB): 54.85 -> 5.47 | Helped |
| Use adaptive anisotropic sub-sampling instead of FineRecon's uniform 1 cm sampling | AniGrad · ScanNet test set | SEL (s): 24.81 -> 1.11 | Helped |
| Use adaptive anisotropic sub-sampling instead of FineRecon's uniform 1 cm sampling | AniGrad · ScanNet test set | Chamfer distance (cm): 4.39 -> 5.29 | Hurt |
| Add basis gradient loss Lbasis (ablation) | AniGrad · ScanNet test set | SEL (s): 2.8 -> 1.11 | Helped |
| Add TSDF gradient loss Lgrad (ablation) | AniGrad · ScanNet test set | F-score: 73.1 -> 74.0 | Helped |
| Add surface TSDF loss Lsurf (ablation) | AniGrad · ScanNet test set | F-score: 70.1 -> 74.0 | Helped |
| Add surface TSDF loss Lsurf and TSDF gradient loss Lgrad (ablation) | AniGrad · ScanNet test set | F-score: 71.4 -> 74.0 | Helped |
Surprising findings
- Applying the same gradient-bound derivation to prior models such as Atlas/VoRTX was found to give bounds too loose to be useful, causing extreme over-sampling.
Limitations
- Resolution is adaptive only within each voxel, so it cannot be coarser than the 4 cm voxel size.
- Sample rates are axis-aligned, so gains are greatest for predominantly axis-aligned surfaces; sliver triangles can result.
- Chamfer distance is second-best rather than best, attributed to not using depth guidance.
Experimental setup
Datasets: ScanNet ScanNet++
Benchmarks: ScanNet
Metrics: F-Score Chamfer distance mesh size surface extraction latency per-frame time
Baselines: Atlas NeuralRecon VoRTX FineRecon
Builds on: FineRecon Atlas Neural Poisson dual marching cubes
Closest papers by meaning
- ND-SDF: Learning Normal Deflection Fields for High-Fidelity Indoor Reconstruction (ICLR 2025 · similarity 0.84)
- SPAR3D: Stable Point-Aware Reconstruction of 3D Objects from Single Images (CVPR 2025 · similarity 0.83)
- Adaptive 3D Reconstruction via Diffusion Priors and Forward Curvature-Matching Likelihood Updates (NeurIPS 2025 · similarity 0.83)
- NC-SDF: Enhancing Indoor Scene Reconstruction Using Neural SDFs with View-Dependent Normal Compensation (CVPR 2024 · similarity 0.83)
- Reconstruct, Inpaint, Test-Time Finetune: Dynamic Novel-view Synthesis from Monocular Videos (NeurIPS 2025 · similarity 0.83)
- Efficient-3Dim: Learning a Generalizable Single-image Novel-view Synthesizer in One Day (ICLR 2024 · similarity 0.83)
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.