Research map / 3D vision / Face and shape
LTM: Lightweight Textured Mesh Extraction and Refinement of Large Unbounded Scenes for Efficient Storage and Real-time Rendering
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.
Produces lightweight textured meshes for large unbounded scenes by baking a BakedSDF-style neural SDF, decimating the mesh with geometry-aware QSlim, and jointly refining vertices and an appearance MLP via differentiable rasterization.
Problem
Baked neural SDF meshes of large unbounded scenes have huge triangle counts and storage footprints, and naive gradient-based mesh optimization with differentiable rendering is unstable for such meshes.
Approach
Combine BakedSDF geometry with region-aware marching cubes and aggressive background decimation (vertex clustering plus normal-aware QSlim), then jointly optimize an appearance MLP and per-vertex deformation with image L1, Laplacian, normal-consistency and normal-map losses, with subdivision-based mesh refinement, so the mesh stays rasterizable with few triangles.
Key results
- On mip-NeRF 360 outdoor scenes, the method is within 0.3 PSNR of BakedSDF (23.2 vs 23.5) while using about 73x fewer triangles and 11x less disk storage.
- On indoor mip-NeRF 360 scenes it reaches 27.1 PSNR vs 27.0 for BakedSDF.
- Compared with NeRF2Mesh and MobileNeRF it gives better PSNR on mip-NeRF 360 and Tanks & Temples and runs at 212 FPS with 115 MB GPU memory on a MacBook Pro M1 (vs 205 FPS / 130 MB for NeRF2Mesh).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the appearance and geometry optimization stage (ablation) | LTM · mip-NeRF 360 outdoor mean | PSNR: 18.1 -> 23.2 | Helped |
| add the appearance and geometry optimization stage (ablation) | LTM · mip-NeRF 360 indoor mean | PSNR: 17.0 -> 27.1 | Helped |
| add vertex deformation training (ablation) | LTM · mip-NeRF 360 outdoor mean | PSNR: 22.5 -> 23.2 | Helped |
| add vertex deformation training (ablation) | LTM · mip-NeRF 360 indoor mean | PSNR: 25.4 -> 27.1 | Helped |
| use rotation-invariant ADAM for vertex updates (ablation) | LTM · mip-NeRF 360 outdoor and indoor means | PSNR: 22.8 vs 23.2; 26.0 vs 27.1 | Hurt |
Surprising findings
- Rasterization with the initial volumetric-trained appearance field and no mesh optimization gives very poor rendering quality (17.0 PSNR mean on mip-NeRF 360), showing the optimization stage is essential.
Limitations
- Appearance model is simpler than BakedSDF's spherical Gaussian model, limiting possible rendering quality.
- Gradient-based vertex optimization may still cause mesh distortions including self-intersections.
- Geometry quality still trails BakedSDF's finer meshes.
Experimental setup
Datasets: mip-NeRF 360 Tanks & Temples Deep Blending
Benchmarks: mip-NeRF 360 Tanks & Temples Deep Blending
Metrics: PSNR SSIM LPIPS disk storage (MB) number of vertices and faces FPS GPU memory
Baselines: MobileNeRF NeRF2Mesh BakedSDF NeRFMeshing Instant-NGP mip-NeRF 360
Builds on: BakedSDF NeRF2Mesh mip-NeRF 360 VolSDF nvdiffrast QSlim
Closest papers by meaning
- S2MDF: A Plug-And-Play Layer for Intersection-Free Multi-Object Signed Distance Fields (NeurIPS 2026 · similarity 0.87)
- High Resolution UDF Meshing via Iterative Networks (NeurIPS 2025 · similarity 0.86)
- Flatten Anything: Unsupervised Neural Surface Parameterization (NeurIPS 2024 · similarity 0.86)
- Neural Signed Distance Function Inference through Splatting 3D Gaussians Pulled on Zero-Level Set (NeurIPS 2024 · similarity 0.86)
- GTR: Improving Large 3D Reconstruction Models through Geometry and Texture Refinement (ICLR 2025 · similarity 0.86)
- DUDF: Differentiable Unsigned Distance Fields with Hyperbolic Scaling (CVPR 2024 · similarity 0.85)
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.