Research map / 3D vision / NeRF and radiance fields
Ced-NeRF: A Compact and Efficient Method for Dynamic 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.
Ced-NeRF builds a fast dynamic NeRF on an Instant-NGP hash-grid backbone with hierarchical motion prediction, neural latent regularization, and time-feature attenuation to curb overfitting and speed up training and rendering.
Problem
Real-time, fast-training dynamic NeRF rendering is hard; combining hybrid voxel/grid representations with a canonical-space deformation field causes overfitting and floater/noisy-geometry artifacts on novel views.
Approach
Adds three components to a grid-based dynamic NeRF: a two-stage (coarse cell-level plus fine within-cell) motion prediction, a training-only small MLP that regresses grid features as a global regularizer, and a time-feature attenuation scaled by deformation norm.
Key results
- On D-NeRF, Ced-NeRF (20K steps) reaches 34.21 PSNR and 0.99 SSIM in 4 minutes of training, beating TiNeuVox (32.67) and TiDNeRF (29.84).
- On HyperNeRF vrig scenes, averages 24.4 PSNR / 0.823 SSIM in 9 minutes training with 0.16 s/img rendering, versus NeRFPlayer at 4.80 s/img.
- On the Plenoptic Video dataset, reaches 30.6 PSNR in 21 minutes, versus 1344 hours for DyNeRF, but lower SSIM (0.919) than K-Planes (0.964).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Ced-NeRF instead of TiNeuVox | dynamic radiance fields · D-NeRF | PSNR: 32.67 -> 34.21 | Helped |
| Use Ced-NeRF instead of TiNeuVox | dynamic radiance fields · D-NeRF | LPIPS: 0.041 -> 0.037 | Helped |
| Use Ced-NeRF instead of TiNeuVox | dynamic radiance fields · D-NeRF | training time: 30 min -> 4 min | Helped |
| Use Ced-NeRF instead of TiNeuVox | dynamic radiance fields · HyperNeRF | mean PSNR: 24.3 -> 24.4 | Helped |
| Use Ced-NeRF instead of TiNeuVox | dynamic radiance fields · HyperNeRF | mean SSIM: 0.837 -> 0.823 | Hurt |
| Use Ced-NeRF instead of K-Planes | dynamic radiance fields · Plenoptic Video | PSNR: 31.6 -> 30.6 | Hurt |
| Use Ced-NeRF instead of K-Planes | dynamic radiance fields · Plenoptic Video | training time: 1.8 hours -> 21 min | Helped |
| Add neural latent regularization (ablation) | Ced-NeRF · D-NeRF | PSNR: +0.25 dB | Helped |
| Add hierarchical motion prediction (ablation) | Ced-NeRF · D-NeRF | PSNR: 33.3 -> 34.0 | Helped |
| Add time feature attenuation (ablation) | Ced-NeRF · D-NeRF | PSNR: 34.0 -> 34.2 | Helped |
Surprising findings
- Neural latent regularization improves PSNR by only about 0.25 dB in ablation, though it visibly reduces artifacts.
- On the Plenoptic Video dataset, SSIM is lower than DyNeRF and K-Planes.
Limitations
- Overfitting gains from latent regularization and time attenuation are small in ablation (about 0.25 dB and 0.2 dB).
- SSIM on the Plenoptic Video dataset trails DyNeRF and K-Planes.
- Hyperparameters (alpha, lambda, xi) are set differently for synthetic and real scenes.
Experimental setup
Datasets: D-NeRF HyperNeRF Plenoptic Video (DyNeRF)
Benchmarks: D-NeRF HyperNeRF Plenoptic Video
Metrics: PSNR SSIM LPIPS training time render time
Baselines: NeRF Nerfies HyperNeRF NeRFPlayer TiNeuVox D-NeRF K-Planes TiDNeRF DyNeRF LLFF
Builds on: Instant-NGP D-NeRF Nerfacc HyperNeRF
Closest papers by meaning
- NeRFLight: Fast and Light Neural Radiance Fields Using a Shared Feature Grid (CVPR 2023 · similarity 0.92)
- Rad-NeRF: Ray-decoupled Training of Neural Radiance Field (NeurIPS 2024 · similarity 0.92)
- NeRF On-the-go: Exploiting Uncertainty for Distractor-free NeRFs in the Wild (CVPR 2024 · similarity 0.90)
- A General Implicit Framework for Fast NeRF Composition and Rendering (AAAI 2024 · similarity 0.90)
- NeRF-VPT: Learning Novel View Representations with Neural Radiance Fields via View Prompt Tuning (AAAI 2024 · similarity 0.90)
- VideoRF: Rendering Dynamic Radiance Fields as 2D Feature Video Streams (CVPR 2024 · similarity 0.89)
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.