Research map / 3D vision / NeRF and radiance fields
BLiRF: Bandlimited Radiance Fields for Dynamic Scene Modeling
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 BLiRF, a dynamic radiance field that treats density and light fields as bandlimited high-dimensional signals, factorizing space and time with low-rank VM-style spatial priors and a learned neural trajectory basis for time.
Problem
Recovering the 3D structure and appearance of non-rigid dynamic scenes from a single moving camera is under-constrained, and existing NeRF-based ray deformation methods suffer from canonical-frame dependence, light/density entanglement, limited expressiveness, and poor motion localization.
Approach
Replace ray deformation with a bandlimited space-time factorization of the radiance field: time is modeled by a learned neural trajectory basis and space by a low-rank vector-matrix (VM) decomposition, plus a locally low-dimensional manifold regularizer based on sinc interpolation and total variation losses.
Key results
- On the custom synthetic 'Color Change' scene, BLiRF reaches 36.68 PSNR / 0.97 SSIM / 0.08 LPIPS versus 29.45 PSNR for the best baseline T-TensoRF (0.88 SSIM, 0.24 LPIPS on Cat, and 35.16 PSNR for T-TensoRF on Color Change).
- On the real-world Cat scene, BLiRF achieves 29.69 PSNR and 0.21 LPIPS compared with 29.45 PSNR for T-TensoRF and 27.49 PSNR for D-NeRF.
- Ablation on time basis shows the learned neural basis outperforms DCT, Fourier and Bernstein bases (e.g., 36.68 vs 33.99 PSNR on Color Change with DCT).
- The method reportedly trains in around 3 hours per scene on a single V100, more than 10 times faster than NR-NeRF, D-NeRF, and HyperNeRF.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use BLiRF instead of T-TensoRF | dynamic-scene radiance-field modeling · Color Change | PSNR: 35.16 -> 36.68 | Helped |
| Use BLiRF instead of D-NeRF | dynamic-scene radiance-field modeling · Falling and Scale | PSNR: 24.60 -> 35.74 | Helped |
| Use BLiRF instead of HyperNeRF | dynamic-scene radiance-field modeling · Fist | PSNR: 30.7 -> 28.4 | Hurt |
| Use BLiRF instead of HyperNeRF | dynamic-scene radiance-field modeling · Teapot | PSNR: 26.4 -> 26.1 | Hurt |
| Replace DCT basis with learned neural trajectory basis (ablation) | BLiRF · Color Change | PSNR: 33.99 -> 36.68 | Helped |
| Replace Fourier basis with learned neural trajectory basis (ablation) | BLiRF · Color Change | PSNR: 31.33 -> 36.68 | Helped |
| Replace Bernstein basis with learned neural trajectory basis (ablation) | BLiRF · Color Change | PSNR: 27.81 -> 36.68 | Helped |
Surprising findings
- Ray deformation networks are shown to model light and density changes separately to some extent, contrary to the assumption that they cannot disentangle them.
Limitations
- Framework is generic but only one example implementation with specific priors is explored.
- Baselines were tuned by grid search per scene while the proposed model uses a single hyperparameter setting, making some comparisons dependent on tuning protocol.
- Factorization in Eq. 10 violates the strict space-time factorization in practice, handled only approximately.
Experimental setup
Datasets: HyperNeRF dataset NVIDIA dynamic scene dataset DyCheck dataset
Benchmarks: HyperNeRF dataset custom synthetic and real-world dynamic scene dataset
Metrics: PSNR SSIM LPIPS
Baselines: D-NeRF NR-NeRF TiNeuVox HyperNeRF TensoRF T-TensoRF Nerfies NV NSFF
Builds on: NeRF TensoRF VM-decomposition D-NeRF HyperNeRF
Closest papers by meaning
- NeRF On-the-go: Exploiting Uncertainty for Distractor-free NeRFs in the Wild (CVPR 2024 · similarity 0.90)
- Joint Optimization of Neural Radiance Fields and Continuous Camera Motion from a Monocular Video (CVPR 2025 · similarity 0.89)
- NeRF-HuGS: Improved Neural Radiance Fields in Non-static Scenes Using Heuristics-Guided Segmentation (CVPR 2024 · similarity 0.89)
- CF-NeRF: Camera Parameter Free Neural Radiance Fields with Incremental Learning (AAAI 2024 · similarity 0.89)
- Ced-NeRF: A Compact and Efficient Method for Dynamic Neural Radiance Fields (AAAI 2024 · similarity 0.89)
- Mask-Based Modeling for Neural Radiance Fields (ICLR 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.