Research map / 3D vision / NeRF and radiance fields
PruNeRF: Segment-Centric Dataset Pruning via 3D Spatial Consistency
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 PruNeRF, a segment-centric dataset pruning method that uses Influence-Function self-influence scores, depth-based 3D reprojection consistency, and SAM-based segment refinement to remove distractors from NeRF training images.
Problem
NeRF training quality degrades when transient distractor objects appear in only some training views, and manually curating datasets to remove unknown distractors is prohibitively expensive.
Approach
Combines per-pixel Self-Influence scores (Influence Functions, approximated with last-layer Hessian) with depth-based reprojection across views to test 3D consistency, then refines pixel masks to segment level using SAM, pruning distractor segments from the training set without a clean validation set.
Key results
- On Kubric synthetic scenes PruNeRF averages about 1.08 dB PSNR above RobustNeRF (e.g., Bag 41.23 vs 39.34).
- On natural scenes PruNeRF reaches 21.31 PSNR on Statue, 22.86 on Android, and 31.73 on BabyYoda, versus RobustNeRF's 20.89, 21.72, and 30.87.
- Ablation on BabyYoda shows Influence Functions alone raise PSNR from 25.22 to 30.40 over the Mip-NeRF 360 baseline, and adding 3D consistency and segmentation reaches 31.73.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use PruNeRF instead of RobustNeRF | NeRF robustness to distractors · Synthetic Kubric scenes | PSNR: 1.08 | Helped |
| use PruNeRF instead of RobustNeRF | NeRF robustness to distractors · Natural scenes | PSNR: 0.8 PSNR | Helped |
| add Influence Functions (ablation) | Mip-NeRF360 · Statue, Android, and BabyYoda | PSNR: 19.64 -> 19.90; 19.53 -> 20.23; 25.22 -> 30.40 | Helped |
| add 3D spatial consistency to Influence Functions (ablation) | Mip-NeRF360 + IF · Statue, Android, and BabyYoda | PSNR: 19.90 -> 20.93; 20.23 -> 20.93; 30.40 -> 31.32 | Helped |
| add pixel-to-segment refinement (ablation) | Mip-NeRF360 + IF + 3D spatial consistency · Statue, Android, and BabyYoda | PSNR: 20.93 -> 21.31; 20.93 -> 21.22; 31.32 -> 31.73 | Helped |
| use Influence Functions instead of loss or gradient-norm (ablation) | PruNeRF distraction measurement · Statue, Android, and BabyYoda, top-5% and top-10% pruning | PSNR: 19.87, 19.76; 19.96, 19.88 -> 19.98, 20.29; 22.20, 22.18; 21.82, 21.78 -> 22.32, 22.42; 30.37, 30.46; 30.68, 30.68 -> 30 | Helped |
| use Arnoldi approximation instead of last-layer Influence Functions (ablation) | PruNeRF · Statue | PSNR: 21.31 -> 21.46 | Helped |
| use Arnoldi approximation instead of last-layer Influence Functions (ablation) | PruNeRF · Android | PSNR: 22.86 -> 22.97 | Helped |
| use Arnoldi approximation instead of last-layer Influence Functions (ablation) | PruNeRF · BabyYoda | PSNR: 31.73 -> 31.93 | Helped |
Surprising findings
- Loss and gradient-norm metrics sometimes fail to separate distractor pixels from hard-to-learn regions, whereas Influence Functions do better.
- D^2NeRF performs worse than plain Mip-NeRF 360 on several natural scenes (e.g., BabyYoda 17.32 PSNR).
Limitations
- Relies on hyperparameters such as the 3-sigma outlier threshold, depth-consistency threshold 0.1, and segment ratio epsilon 0.1, set fixed across datasets.
- Uses only the last layer for Hessian approximation, which may be inaccurate for NeRF; Arnoldi approximation is an extra variant.
- Evaluated on few scenes; sparse-view and end-to-end optimization left to future work.
Experimental setup
Datasets: Kubric Pick Statue Android BabyYoda
Benchmarks: Kubric Statue Android BabyYoda Pick
Metrics: PSNR SSIM MS-SSIM
Baselines: NeRF-W NSFF NeuralDiff D^2NeRF RobustNeRF Mip-NeRF 360 (l1, l2, Charbonnier)
Builds on: NeRF Mip-NeRF 360 Influence Functions RobustNeRF SAM
Closest papers by meaning
- NeRF-HuGS: Improved Neural Radiance Fields in Non-static Scenes Using Heuristics-Guided Segmentation (CVPR 2024 · similarity 0.88)
- Feature 3DGS: Supercharging 3D Gaussian Splatting to Enable Distilled Feature Fields (CVPR 2024 · similarity 0.87)
- NeRF On-the-go: Exploiting Uncertainty for Distractor-free NeRFs in the Wild (CVPR 2024 · similarity 0.86)
- VLM-Pruner: Buffering for Spatial Sparsity in an Efficient VLM Centrifugal Token Pruning Paradigm (CVPR 2026 · similarity 0.85)
- NeRFPrior: Learning Neural Radiance Field as a Prior for Indoor Scene Reconstruction (CVPR 2025 · similarity 0.85)
- 3D Gaussian Rendering Can Be Sparser: Efficient Rendering via Learned Fragment Pruning (NeurIPS 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.