Research map / Video / Video segmentation and tracking
TeTriRF: Temporal Tri-Plane Radiance Fields for Efficient Free-Viewpoint Video
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 TeTriRF, a hybrid tri-plane plus density-grid representation for free-viewpoint video, trained in temporal groups with consistency regularizers and compressed with off-the-shelf HEVC video codecs.
Problem
Free-viewpoint video with NeRF requires large storage for long multi-view sequences, and existing compact methods are either slow to render or need high-end GPUs for decoding.
Approach
Factorize each frame into a tri-plane of feature planes and an explicit density grid, train consecutive frame groups jointly with intra- and inter-group L1 regularizers that make features temporally consistent and low-entropy, then compress the planes as image sequences with H.265.
Key results
- On the NHR dataset, Ours(low) reaches 30.42 PSNR at about 11.76 KB per frame, versus ReRF's 30.34 PSNR at 1220 KB, and renders at 0.1 s per frame versus 0.21 s for ReRF.
- On the ReRF dataset, Ours(low) achieves 27.60 PSNR at 11.72 KB per frame, while Ours(high) reaches 30.18 PSNR at 71.67 KB per frame, compared with ReRF's 30.33 PSNR at 843 KB.
- On DyNeRF, Ours(high) reports 30.43 PSNR, comparable to other methods, with lower training time and storage than the alternatives tested (e.g. 1733 KB for MixVoxels).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add two-pass progressive scaling (ablation) | TeTriRF · sport1 (NHR) | PSNR | Helped |
| add group-based regularizers (ablation) | TeTriRF · sport1 (NHR) | PSNR | Helped |
| replace H.265 with MPEG-2 codec (ablation) | TeTriRF · sport1 (NHR) | PSNR | Hurt |
| use the hybrid tri-plane and density-grid representation instead of ReRF’s 3D voxel grid (ablation) | TeTriRF · sport1 (NHR) | PSNR | Helped |
Surprising findings
- Even with the older MPEG-2 codec in place of H.265, TeTriRF still outperforms ReRF in the rate-distortion comparison, suggesting the hybrid representation itself helps compression.
Limitations
- Rendering speed depends on scene sparsity and the Python implementation includes system overhead, so timings are only relative.
- Experiments use stationary cameras; moving-camera dynamic scenes are not covered.
- On the DyNeRF forward-facing scene, Ours(low) has notably lower quality (28.71 PSNR) than the baselines.
Experimental setup
Datasets: NHR ReRF DyNeRF
Benchmarks: NHR ReRF dataset DyNeRF
Metrics: PSNR SSIM LPIPS training time per frame rendering time per frame model size per frame (KB)
Baselines: KPlanes HumanRF TiNeuVox ReRF NeRFPlayer MixVoxels
Builds on: DVGO K-Planes ReRF NeRF deferred shading
Closest papers by meaning
- VideoRF: Rendering Dynamic Radiance Fields as 2D Feature Video Streams (CVPR 2024 · similarity 0.90)
- NeRF-VPT: Learning Novel View Representations with Neural Radiance Fields via View Prompt Tuning (AAAI 2024 · similarity 0.89)
- MuRF: Multi-Baseline Radiance Fields (CVPR 2024 · similarity 0.88)
- HyRF: Hybrid Radiance Fields for Memory-efficient and High-quality Novel View Synthesis (NeurIPS 2025 · similarity 0.88)
- Ced-NeRF: A Compact and Efficient Method for Dynamic Neural Radiance Fields (AAAI 2024 · similarity 0.88)
- Rad-NeRF: Ray-decoupled Training of Neural Radiance Field (NeurIPS 2024 · similarity 0.88)
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.