Research map / 3D vision / Gaussian splatting
Enhancing Close-up Novel View Synthesis via Pseudo-labeling
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 pseudo-label-based training for radiance fields (NeRF and 2DGS) that warps training images to random close-up camera poses, filters them with a consistency mask, and adds a new close-up novel-view benchmark.
Problem
Novel view synthesis from NeRF and Gaussian Splatting methods degrades badly for close-up viewpoints that deviate strongly from the training views, because training data lacks close-up coverage.
Approach
Generate random close-up virtual camera poses around training anchor points, warp training images to them using rendered depth, and keep only pixels whose warped values agree with a multi-view occlusion-aware warp (mask), using these as pseudo-labels for fine-tuning; additionally allow test-time fine-tuning on known test poses.
Key results
- On the proposed close-up dataset, Ours-NeRF reaches PSNR 18.92 vs. 14.48 for NeRF (+4.44).
- Ours-2DGS reaches PSNR 20.88 / SSIM 0.784 / LPIPS 0.383, versus 19.99 PSNR for 2DGS and 19.16 for 3DGS.
- Ablation: pseudo-label mask raises PSNR from 16.61 to 18.17; random close-up pose generation adds to 18.92.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply pseudo-label fine-tuning to 2DGS | 2DGS · Close-up dataset | PSNR: 19.99 -> 20.88 | Helped |
| Apply pseudo-label fine-tuning to NeRF | NeRF · Close-up dataset | PSNR: 14.48 -> 18.92 | Helped |
| Add initial warped pseudo-labels (ablation) | NeRF fine-tuning · Close-up dataset | PSNR: 14.48 -> 16.61 | Helped |
| Add a mask to filter erroneous pseudo-label regions (ablation) | NeRF fine-tuning with initial pseudo-labels · Close-up dataset | PSNR: 16.61 -> 18.17 | Helped |
| Add random diverse close-up perspective generation (ablation) | Masked pseudo-label fine-tuning · Close-up dataset | PSNR: 18.17 -> 18.92 | Helped |
| Apply test-time pseudo-label fine-tuning (ablation) | 2DGS · Close-up dataset, test-time fine-tuning | PSNR: 19.99 -> 20.95 | Helped |
Surprising findings
- Mip-Splatting, designed for alias-free rendering, scores below baseline 2DGS on close-up views (19.55 vs 19.99 PSNR).
Limitations
- Pseudo-labels from warped training images can contain errors, which requires masking and may leave some regions unsupervised.
- Results are reported only on the authors' own new dataset, with no evaluation on other close-up benchmarks.
Experimental setup
Datasets: LLFF Tanks & Temples
Benchmarks: Diverse close-up perspectives dataset (proposed)
Metrics: PSNR SSIM LPIPS
Baselines: NeRF 2DGS 3DGS Mip-NeRF Mip-Splatting Instant-NGP TensoRF Zip-NeRF DS-NeRF
Builds on: NeRF 2D Gaussian Splatting 3D Gaussian Splatting
Closest papers by meaning
- FewViewGS: Gaussian Splatting with Few View Matching and Multi-stage Training (NeurIPS 2024 · similarity 0.88)
- MACRO: Training-free Multi-plane Attention for Closeup Render Optimization (NeurIPS 2026 · similarity 0.86)
- Feature 3DGS: Supercharging 3D Gaussian Splatting to Enable Distilled Feature Fields (CVPR 2024 · similarity 0.86)
- NerfBaselines: Consistent and Reproducible Evaluation of Novel View Synthesis Methods (NeurIPS 2025 · similarity 0.86)
- Structure Consistent Gaussian Splatting with Matching Prior for Few-shot Novel View Synthesis (NeurIPS 2024 · similarity 0.86)
- NeRF-VPT: Learning Novel View Representations with Neural Radiance Fields via View Prompt Tuning (AAAI 2024 · similarity 0.86)
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.