Research map / 3D vision / NeRF and radiance fields
L0-Sampler: An L0 Model Guided Volume Sampling for NeRF
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 L0-Sampler, which replaces NeRF's piecewise-constant weight interpolation in hierarchical volume sampling with piecewise exponential interpolation to get a quasi-L0 weight function for parameter-free, few-line sampling improvements.
Problem
Hierarchical volume sampling (HVS) in NeRF approximates the volume rendering weight function with piecewise constant functions, giving a coarse estimate that spreads fine samples uniformly within intervals rather than concentrating them near the surface.
Approach
Observes that the well-trained weight function w(t) resembles the L0 distance to the surface, and interpolates coarse weights with piecewise exponential functions a(b/a)^s (instead of constants) so that the normalized PDF is steep near high-weight endpoints, concentrating fine samples near surfaces without extra parameters or loss terms.
Key results
- Applying L0-Sampler to NeRF improves Blender PSNR, e.g. Lego 31.39 to 31.97 and Chair 34.52 to 34.92, at nearly the same training time.
- On NeuS with DTU, mean Chamfer distance drops from 0.848 to 0.810.
- Sparsity loss gives smaller gains than L0-Sampler on Blender (e.g. Lego PSNR 31.58 vs 31.97 for L0-Sampler).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace HVS with L0-Sampler | NeRF · Blender scenes | PSNR: 31.39 -> 31.97; 34.52 -> 34.92; 25.59 -> 25.72; 29.47 -> 29.80; 28.92 -> 29.21; 29.59 -> 29.77; 36.80 -> 37.02; 33.18 -> | Helped |
| Replace HVS with L0-Sampler | NeuS · DTU mean | Chamfer distance: 0.848 -> 0.810 | Helped |
| Replace HVS with L0-Sampler | NeRF · Blender scenes | LPIPS: 0.0400 -> 0.0346; 0.0283 -> 0.0257; 0.0741 -> 0.0685; 0.1409 -> 0.1342; 0.0416 -> 0.0368; 0.0432 -> 0.0386; 0.0298 -> 0. | Helped |
| Use piecewise exponential rather than constant interpolation (ablation) | Instant NGP · Blender Chair | PSNR: 32.07 -> 33.05 | Helped |
| Use piecewise exponential rather than constant interpolation (ablation) | Instant NGP · LLFF Fern | PSNR: 26.33 -> 26.56 | Helped |
| Use cubic spline interpolation instead of piecewise exponential interpolation (ablation) | Instant NGP · Blender Chair | PSNR: 33.05 -> 33.01 | Hurt |
| Use Akima interpolation instead of piecewise exponential interpolation (ablation) | Instant NGP · Blender Chair | PSNR: 33.05 -> 32.96 | Hurt |
| Use inverse interpolation instead of piecewise exponential interpolation (ablation) | Instant NGP · Blender Chair | PSNR: 33.05 -> 32.54 | Hurt |
| Use linear interpolation instead of piecewise exponential interpolation (ablation) | Instant NGP · Blender Chair | PSNR: 33.05 -> 32.86 | Hurt |
| Use inverse interpolation instead of piecewise exponential interpolation (ablation) | Instant NGP · LLFF Fern | PSNR: 26.56 -> 26.54 | Hurt |
| Use linear interpolation instead of piecewise exponential interpolation (ablation) | Instant NGP · LLFF Fern | PSNR: 26.56 -> 26.41 | Hurt |
Surprising findings
- Piecewise inverse interpolation performs less consistently than the exponential one despite similar bias.
- Cubic spline and Akima interpolation, though smooth and widely used, do not match the gains of the piecewise exponential function.
Limitations
- Not all results improve; gains are often modest.
- Method only modifies the fine-stage sampling; applying the L0 model to the coarse stage is left open.
Experimental setup
Datasets: NeRF Blender synthetic LLFF LF dataset DTU BlendedMVS bunny smoke fox
Benchmarks: NeRF Blender synthetic LLFF LF dataset DTU BlendedMVS
Metrics: PSNR SSIM LPIPS Chamfer distance
Baselines: NeRF mip-NeRF Instant NGP NeRF++ NeuS sparsity loss piecewise linear interpolation cubic spline Akima
Builds on: NeRF hierarchical volume sampling mip-NeRF Instant NGP NeuS
Closest papers by meaning
- GL-NeRF: Gauss-Laguerre Quadrature Enables Training-Free NeRF Acceleration (NeurIPS 2024 · similarity 0.88)
- CVT-xRF: Contrastive In-Voxel Transformer for 3D Consistent Radiance Fields from Sparse Inputs (CVPR 2024 · similarity 0.85)
- Neural NeRF Compression (ICML 2024 · similarity 0.85)
- Rad-NeRF: Ray-decoupled Training of Neural Radiance Field (NeurIPS 2024 · similarity 0.85)
- NeRFPrior: Learning Neural Radiance Field as a Prior for Indoor Scene Reconstruction (CVPR 2025 · similarity 0.85)
- NerfBaselines: Consistent and Reproducible Evaluation of Novel View Synthesis Methods (NeurIPS 2025 · similarity 0.84)
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.