Research map / 3D vision / NeRF and radiance fields
DaReNeRF: Direction-aware Representation for Dynamic Scenes
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 DaReNeRF, a NeRF-based dynamic-scene representation that replaces plane-based features with direction-aware dual-tree complex wavelet coefficients (six real and six imaginary orientations), plus learned masks for compression.
Problem
Plane-based explicit decompositions for dynamic NeRF train fast but lose high-frequency texture and struggle with complex motion; plain DWT wavelet representations suffer shift variance and poor direction selectivity.
Approach
Builds each 4D plane-based feature from an inverse dual-tree complex wavelet transform over six directional real and imaginary coefficient maps, giving shift invariance and orientation selectivity, and uses trainable masks with a Huffman/RLE compression pipeline to offset the storage redundancy.
Key results
- On the Plenoptic video dataset (all-scene average), DaReNeRF reaches 32.258 PSNR versus 31.730 for Mix Voxels-X and 31.705 for HexPlane (650k steps), with 4.5h training and 1,210 MB model size.
- On NSVF static scenes, the DTCWT-based representation reaches 36.24 PSNR versus 34.67 for the DWT-based masked wavelet NeRF at similar sparsity (94.2%, 8.98 MB).
- Replacing DWT with the direction-aware DTCWT representation improves PSNR by 0.47, 1.57, and 0.60 over DWT on the three static datasets at similar sparsity.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use the direction-aware representation instead of the plane-based baseline | DaReNeRF on Plenoptic Video, flame-salmon scene · Plenoptic Video, flame-salmon scene | PSNR: 29.581 -> 30.441 | Helped |
| use the direction-aware representation instead of the strongest listed baseline | DaReNeRF on Plenoptic Video, all scenes (average) · Plenoptic Video, all scenes (average) | PSNR: 31.730 -> 32.258 | Helped |
| use DaReNeRF without a deformation field rather than 4D-GS | D-NeRF · D-NeRF | PSNR: 33.30 -> 31.95 | Hurt |
| apply the direction-aware representation to VM-192 (300) | VM-192 (300) on NeRF synthetic · NeRF synthetic | PSNR: 32.91 -> 32.42 | Hurt |
| apply the direction-aware representation to VM-192 (300) | VM-192 (300) on NSVF · NSVF | PSNR: 36.11 -> 36.24 | Helped |
| apply the direction-aware representation to VM-96 (640) | VM-96 (640) on LLFF · LLFF | PSNR: 26.66 -> 26.48 | Hurt |
| use DTCWT with Near Symmetric A instead of DWT with biorthogonal 4.4 (ablation) | DaReNeRF on NSVF · NSVF | PSNR: 34.67 -> 36.24 | Helped |
| use DTCWT with Antonini instead of DWT with biorthogonal 4.4 (ablation) | DaReNeRF on NSVF · NSVF | PSNR: 34.67 -> 36.10 | Helped |
| use DTCWT with LeGall instead of DWT with biorthogonal 4.4 (ablation) | DaReNeRF on NSVF · NSVF | PSNR: 34.67 -> 36.14 | Helped |
| use DTCWT with Near Symmetric B instead of DWT with biorthogonal 4.4 (ablation) | DaReNeRF on NSVF · NSVF | PSNR: 34.67 -> 36.17 | Helped |
| vary the sparsity weight λm from 0 to 2.5 × 10−11 (ablation) | DaReNeRF on NSVF · NSVF | PSNR: 36.34 -> 36.24 | Hurt |
| increase the wavelet level (ablation) | DaReNeRF · scene reconstruction | PSNR | No clear effect |
Surprising findings
- The DWT-based plane representation degrades substantially when applied to dynamic scenes, as reported by HexPlane and confirmed here.
- Increasing the wavelet level beyond one did not significantly improve quality while greatly increasing training time and model size.
Limitations
- Performs poorly under extremely sparse observations such as the D-NeRF monocular dataset, since no deformation field is used.
- Lower compactness than DWT representation; cannot reach sub-1MB model sizes on static scenes.
- Storage redundancy of the 2d factor requires masking and compression to be practical.
Experimental setup
Datasets: Plenoptic Video Dataset (DyNeRF) D-NeRF dataset NeRF synthetic NSVF LLFF
Benchmarks: Plenoptic Video Dataset D-NeRF dataset NeRF synthetic NSVF LLFF
Metrics: PSNR D-SSIM LPIPS SSIM training time model size
Baselines: DyNeRF HexPlane K-Planes Mix Voxels 4D-GS NeRFPlayer HyperReel D-NeRF TiNeuVox masked wavelet NeRF (DWT)
Builds on: HexPlane TensoRF masked wavelet NeRF DyNeRF dual-tree complex wavelet transform
Closest papers by meaning
- 4D3R: Motion-Aware Neural Reconstruction and Rendering of Dynamic Scenes from Monocular Videos (NeurIPS 2025 · similarity 0.88)
- Real-time Photorealistic Dynamic Scene Representation and Rendering with 4D Gaussian Splatting (ICLR 2024 · similarity 0.87)
- TeTriRF: Temporal Tri-Plane Radiance Fields for Efficient Free-Viewpoint Video (CVPR 2024 · similarity 0.86)
- EmerNeRF: Emergent Spatial-Temporal Scene Decomposition via Self-Supervision (ICLR 2024 · similarity 0.86)
- CodecNeRF: Toward Fast Encoding and Decoding, Compact, and High-quality Novel-view Synthesis (AAAI 2025 · similarity 0.86)
- BLiRF: Bandlimited Radiance Fields for Dynamic Scene Modeling (AAAI 2024 · similarity 0.85)
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.