Research map / 3D vision / NeRF and radiance fields
Flow-NeRF: Joint Learning of Geometry, Poses, and Dense Flow within Unified Neural Representations
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.
Flow-NeRF jointly learns scene geometry, camera poses, and dense optical flow in one unposed NeRF, using a pose-conditioned bijective (RealNVP) flow branch and feature message passing to improve novel-view synthesis and depth.
Problem
Pose-free neural radiance field reconstruction suffers from geometric ambiguity and lacks long-range cross-frame consistency; flow is typically used only as a pose regularizer and novel-view dense flow is not inferred.
Approach
Two branches (geometry and flow) share sampled points and a canonical space; a RealNVP bijective mapping conditioned on learnable camera pose predicts dense flow for novel views, and canonical-space features are concatenated into the geometry MLP to enhance it.
Key results
- Ballroom novel-view PSNR rises to 28.83 with flow enhancement versus 24.85 without and 25.33 for Nope-NeRF.
- On Sintel at frame interval 16, average EPE of 1.683 versus RAFT's 2.089.
- ScanNet depth Abs Rel of 0.047 versus 0.141 for Nope-NeRF; mean ScanNet novel-view PSNR 32.55 vs 31.86 for Nope-NeRF.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| jointly learn scene geometry, camera poses, and dense flow with Flow-NeRF | pose-free NeRF · ScanNet | PSNR: 31.86 -> 32.55 | Helped |
| jointly learn scene geometry, camera poses, and dense flow with Flow-NeRF | pose-free NeRF · Tanks and Temples | PSNR: 26.34 -> 28.73 | Helped |
| jointly learn scene geometry, camera poses, and dense flow with Flow-NeRF | pose-free NeRF · ScanNet | Abs Rel: 0.141 -> 0.047 | Helped |
| jointly learn scene geometry, camera poses, and dense flow with Flow-NeRF | pose-free NeRF · Sintel mountain1, frame interval 16 | EPE: 2.089 -> 1.683 | Helped |
| add flow feature message passing (ablation) | Flow-NeRF · Ballroom | PSNR: 24.85 -> 28.83 | Helped |
| add flow feature message passing (ablation) | Flow-NeRF · ScanNet 0418 00 | Abs Rel: 0.200 -> 0.034 | Helped |
| use perspective projection instead of orthogonal projection for flow-branch point sampling (ablation) | Flow-NeRF with flow enhancement · ScanNet 0418 00 | Abs Rel: 0.047 -> 0.034 | Helped |
Surprising findings
- Flow enhancement without perspective projection (orthogonal sampling) yields weaker depth than perspective back-projection, showing sampling geometry matters for flow-to-geometry transfer.
Limitations
- Assumes scenes are nearly static so flow is induced only by camera motion.
- Flow supervision relies on RAFT pseudo flow, so accuracy is bounded by the teacher in training.
Experimental setup
Datasets: Tanks and Temples ScanNet Sintel
Benchmarks: Tanks and Temples ScanNet Sintel
Metrics: PSNR SSIM LPIPS ATE RPEt RPEr Abs Rel Sq Rel RMSE EPE
Baselines: BARF NeRFmm SC-NeRF Nope-NeRF RAFT Omnimotion
Builds on: NeRF Omnimotion Nope-NeRF RealNVP RAFT
Closest papers by meaning
- Bidirectional Optical Flow NeRF: High Accuracy and High Quality under Fewer Views (AAAI 2023 · similarity 0.92)
- Joint Optimization of Neural Radiance Fields and Continuous Camera Motion from a Monocular Video (CVPR 2025 · similarity 0.89)
- Zero-Shot Monocular Scene Flow Estimation in the Wild (CVPR 2025 · similarity 0.88)
- E-MoFlow: Learning Egomotion and Optical Flow from Event Data via Implicit Regularization (NeurIPS 2025 · similarity 0.87)
- DeltaFlow: An Efficient Multi-frame Scene Flow Estimation Method (NeurIPS 2025 · similarity 0.87)
- CF-NeRF: Camera Parameter Free Neural Radiance Fields with Incremental Learning (AAAI 2024 · similarity 0.87)
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.