Research map / 3D vision / NeRF and radiance fields
Gated Fields: Learning Scene Reconstruction from Gated Videos
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.
Gated Fields reconstructs outdoor 3D scene geometry and radiance from active time-gated NIR video by fitting a neural field with a differentiable gated image formation model.
Problem
Outdoor 3D scene reconstruction from RGB fails in low light and texture-poor regions, while scanning LiDAR has low angular resolution and is slow to aggregate for large scenes.
Approach
Integrate a learnable gated imaging formation model (range intensity profiles, pulsed illumination, shadows) into neural volume rendering so that active gated video jointly supervises geometry, ambient light, reflectance and gating/illumination parameters.
Key results
- Depth synthesis: 21.87% lower MAE than the best neural field method and 59.8% lower than the best RGB+LiDAR method, per the abstract; night MAE 4.13 m vs 4.72 m for LiDAR-NeRF.
- 3D reconstruction IoU of 22.25% vs 20.03% for LiDAR-NeRF (about 11% relative gain).
- Novel view synthesis PSNR of 32.28 dB on day data; Mip-NeRF 360 reaches 17.16 dB.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use gated image formation with separate ambient and reflection components instead of a single end-to-end intensity field (ablation) | Gated Fields · Test Data – Night | MAE [m]: 19.95 -> 8.35 | Helped |
| Add depth supervision (ablation) | Gated Fields · Test Data – Night | MAE [m]: 8.35 -> 5.43 | Helped |
| Add depth supervision (ablation) | Gated Fields · Test Data – Day | PSNR [dB]: 26.77 -> 26.88 | Helped |
| Add angular-dependent attenuation and reflectance regularization (ablation) | Gated Fields · Test Data – Night | MAE [m]: 5.43 -> 2.82 | Helped |
| Add angular-dependent attenuation and reflectance regularization (ablation) | Gated Fields · Test Data – Day | MAE [m]: 4.22 -> 4.31 | Hurt |
| Add explicit shadow modeling (ablation) | Gated Fields · Test Data – Night | MAE [m]: 2.82 -> 2.51 | Helped |
| Add explicit shadow modeling (ablation) | Gated Fields · Test Data – Day | MAE [m]: 4.31 -> 4.12 | Helped |
| Use Gated Fields rather than LiDAR-NeRF | LiDAR-NeRF · Test Data – Night | MAE [m]: 4.72 -> 4.13 | Helped |
| Use Gated Fields rather than GatedStereo | GatedStereo · Test Data – Day | MAE [m]: 3.30 -> 2.91 | Helped |
Surprising findings
- Applying state-of-the-art neural field methods directly to gated captures performs poorly, because they cannot model gated imaging formation and only fit the ambient component.
Limitations
- Requires specialized gated camera and VCSEL illuminator hardware.
- Static scenes only; dynamic feedback and adaptive acquisition left to future work.
Experimental setup
Datasets: Gated Fields dataset (collected by authors)
Benchmarks: Gated Fields dataset (day and night test splits)
Metrics: RMSE MAE ARD delta accuracy (1.25^i) PSNR SSIM IoU Precision Recall
Baselines: Gated2Gated Gated Stereo SimIPU AdaBins DPT DepthFormer CREStereo NLSPN Mip-NeRF 360 K-Planes RawNeRF Depth-NeRF
Builds on: Mip-NeRF 360 Gated Stereo hash encoding (Instant-NGP) NeuS-style normal consistency proposal network sampling
Closest papers by meaning
- Flux4D: Flow-based Unsupervised 4D Reconstruction (NeurIPS 2025 · similarity 0.87)
- Neural Inverse Rendering from Propagating Light (CVPR 2025 · similarity 0.86)
- TurboSL: Dense Accurate and Fast 3D by Neural Inverse Structured Light (CVPR 2024 · similarity 0.86)
- PlatoNeRF: 3D Reconstruction in Plato's Cave via Single-View Two-Bounce Lidar (CVPR 2024 · similarity 0.86)
- MorpheuS: Neural Dynamic 360deg Surface Reconstruction from Monocular RGB-D Video (CVPR 2024 · similarity 0.86)
- UniSDF: Unifying Neural Representations for High-Fidelity 3D Reconstruction of Complex Scenes with Reflections (NeurIPS 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.