Research map / 3D vision / NeRF and radiance fields
Depth-Guided Bundle Sampling for Efficient Generalizable Neural Radiance Field 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.
Proposes depth-guided bundle sampling that groups adjacent rays into cones sampled with inscribed spheres, with depth-adaptive sample counts, to speed up generalizable NeRF and 3D-GS rendering.
Problem
Rendering high-resolution novel views with generalizable NeRFs is computationally expensive because every ray is densely sampled, limiting real-time synthesis.
Approach
Group K×K adjacent rays into a bundle sampled jointly via sphere-based cones with mipmapped source-view features (joint bundle plus ray-specific encoding), and allocate samples per bundle adaptively from predicted depth range and confidence, grounded in plenoptic sampling theory.
Key results
- ENeRF+Ours (2x2) improves PSNR by 1.27 dB and FPS by 47% over ENeRF on DTU 3-view (28.86 vs 27.61 PSNR, 28.6 vs 19.5 FPS).
- ENeRF+Ours (4x4) reaches 43.6 FPS on DTU, over twice ENeRF's speed, with 28.21 PSNR.
- On cross-dataset evaluation, ENeRF+Ours (2x2) reaches 24.33 PSNR on Real Forward-facing and 26.49 on NeRF Synthetic (3-view).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use depth-guided bundle sampling with 2 × 2 bundles instead of ENeRF ray sampling | ENeRF · DTU test set, 3-view | PSNR: 27.61 -> 28.86 | Helped |
| Use depth-guided bundle sampling with 2 × 2 bundles instead of ENeRF ray sampling | ENeRF · DTU test set, 3-view | FPS: 19.5 -> 28.6 | Helped |
| Increase bundle size from 2 × 2 to 4 × 4 | ENeRF+Ours · DTU test set, 3-view | FPS: 28.6 -> 43.6 | Helped |
| Increase bundle size from 2 × 2 to 4 × 4 | ENeRF+Ours · DTU test set, 3-view | PSNR: 28.86 -> 28.21 | Hurt |
| Use pixel-aligned feature extraction rather than sphere-based sampling (ablation) | ENeRF+Ours · DTU test set, 3-view, 2 × 2 bundles | PSNR: 28.86 -> 27.66 | Hurt |
| Add depth-guided adaptive sampling rather than fixed sampling (ablation) | ENeRF+Ours · DTU test set, 3-view, 2 × 2 bundles | FPS: 17.0 -> 28.6 | Helped |
| Add ray-specific representation (ablation) | ENeRF+Ours · DTU test set, 3-view, 2 × 2 bundles | PSNR: 28.47 -> 28.86 | Helped |
| Add ray-specific representation (ablation) | ENeRF+Ours · DTU test set, 3-view, 4 × 4 bundles | PSNR: 26.74 -> 28.21 | Helped |
| Add joint bundle representation (ablation) | ENeRF+Ours · DTU test set, 3-view, 2 × 2 bundles | PSNR: 27.83 -> 28.86 | Helped |
| Apply depth-guided bundle sampling to MVSGaussian | MVSGaussian · DTU test set, 3-view | FPS: 21.5 -> 23.4 | Helped |
Surprising findings
- Ray-specific representation adds 0.39 dB at 2x2 bundles and 1.47 dB at 4x4, and substituting sphere-based sampling with pixel-aligned features drops PSNR by 1.2 dB.
Limitations
- MVSGaussian+Ours gains are limited due to fewer sampling points and dependence on post-processing layers.
- Larger bundles (4x4) lose some quality, and the joint bundle representation inherently lacks high-frequency detail.
Experimental setup
Datasets: DTU training set
Benchmarks: DTU Real Forward-facing NeRF Synthetic
Metrics: PSNR SSIM LPIPS FPS average samples per ray
Baselines: PixelNeRF IBRNet MVSNeRF ENeRF MatchNeRF GNT CG-NeRF MuRF ConvGLR MVSGaussian
Builds on: ENeRF MVSGaussian Mip-NeRF plenoptic sampling theory
Closest papers by meaning
- Sparse Voxels Rasterization: Real-time High-fidelity Radiance Field Rendering (CVPR 2025 · similarity 0.86)
- HyRF: Hybrid Radiance Fields for Memory-efficient and High-quality Novel View Synthesis (NeurIPS 2025 · similarity 0.86)
- Learning Robust Generalizable Radiance Field with Visibility and Feature Augmented Point Representation (ICLR 2024 · similarity 0.86)
- HybridNeRF: Efficient Neural Rendering via Adaptive Volumetric Surfaces (CVPR 2024 · similarity 0.85)
- ProbeSDF: Light Field Probes For Neural Surface Reconstruction (CVPR 2025 · similarity 0.85)
- Gaussian Herding across Pens: An Optimal Transport Perspective on Global Gaussian Reduction for 3DGS (NeurIPS 2025 · 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.