Research map / 3D vision / NeRF and radiance fields
Entangled View-Epipolar Information Aggregation for Generalizable Neural Radiance Fields
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 EVE-NeRF, a generalizable NeRF that alternates view and epipolar transformers in dual-branch modules so cross-view and along-epipolar information interact in an entangled way, trained only with photometric loss.
Problem
Generalizable NeRF methods aggregate source-view features either cross-view or along-epipolar, typically independently or sequentially, which causes rendering artifacts and depth discontinuities and limits generalization to new scenes.
Approach
Entangles view and epipolar aggregation via alternating VEI and EVI modules, each with a dual branch: a transformer plus a prior-injecting branch (along-epipolar adjacent-depth attention for appearance continuity; multi-view cross-attention calibration for geometry consistency), using only photometric supervision.
Key results
- On LLFF, EVE-NeRF reaches 27.16 PSNR vs GNT's 25.53 in generalizable NVS setting 1.
- On Shiny, EVE-NeRF reaches 28.01 PSNR and 0.935 SSIM vs GNT's 27.19 and 0.912.
- In few-shot setting (3 views), EVE-NeRF achieves 27.80 PSNR on DTU vs MatchNeRF's 26.91.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use entangled view-epipolar aggregation instead of GNT | EVE-NeRF · LLFF | PSNR: 25.53 -> 27.16 | Helped |
| Use entangled view-epipolar aggregation instead of GPNR | EVE-NeRF · Blender | PSNR: 28.29 -> 27.03 | Hurt |
| Use entangled view-epipolar aggregation instead of GNT | EVE-NeRF · Shiny | PSNR: 27.19 -> 28.01 | Helped |
| Add along-epipolar perception to the view transformer (ablation) | EVE-NeRF · LLFF low-resolution | PSNR: 25.03 -> 25.48 | Helped |
| Add multi-view calibration to the epipolar transformer (ablation) | EVE-NeRF · LLFF low-resolution | PSNR: 25.02 -> 25.17 | Helped |
| Replace the naïve dual-transformer architecture with cross-attention interaction (ablation) | EVE-NeRF · LLFF low-resolution | PSNR: 25.66 -> 25.85 | Helped |
| Use entangled EVE-NeRF aggregation rather than a naïve dual-transformer architecture (ablation) | EVE-NeRF · LLFF low-resolution | PSNR: 25.66 -> 26.69 | Helped |
| Use entangled EVE-NeRF aggregation rather than dual-transformer cross-attention interaction (ablation) | EVE-NeRF · LLFF low-resolution | PSNR: 25.85 -> 26.69 | Helped |
| Use EVE-NeRF rather than MatchNeRF | EVE-NeRF · DTU test | PSNR: 26.91 -> 27.80 | Helped |
| Use EVE-NeRF rather than MatchNeRF | EVE-NeRF · Blender few-shot | LPIPS: 0.164 -> 0.132 | Helped |
Limitations
- Ablations run at reduced resolution (504x378) on LLFF only
- Gains on Blender in few-shot setting are small (23.45 vs MatchNeRF 23.20 PSNR, lower than other metrics)
Experimental setup
Datasets: Google Scanned Objects RealEstate10K Spaces LLFF handheld captures DTU
Benchmarks: LLFF Blender Shiny DTU
Metrics: PSNR SSIM LPIPS
Baselines: PixelNeRF IBRNet GPNR NeuRay GNT MVSNeRF MatchNeRF
Builds on: NeRF IBRNet GNT GPNR PixelNeRF
Closest papers by meaning
- GoLF-NRT: Integrating Global Context and Local Geometry for Few-Shot View Synthesis (CVPR 2025 · similarity 0.85)
- NexusGS: Sparse View Synthesis with Epipolar Depth Priors in 3D Gaussian Splatting (CVPR 2025 · similarity 0.84)
- Epipolar-Free 3D Gaussian Splatting for Generalizable Novel View Synthesis (NeurIPS 2024 · similarity 0.84)
- NeRFPrior: Learning Neural Radiance Field as a Prior for Indoor Scene Reconstruction (CVPR 2025 · similarity 0.84)
- Mask-Based Modeling for Neural Radiance Fields (ICLR 2024 · similarity 0.84)
- ColNeRF: Collaboration for Generalizable Sparse Input Neural Radiance Field (AAAI 2024 · 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.