Research map / Video / Video segmentation and tracking
DS-NeRV: Implicit Neural Video Representation with Decomposed Static and Dynamic Codes
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 DS-NeRV, an implicit neural video representation that decomposes a video into sparse learnable static codes (weighted-sum sampled) and dynamic codes (interpolation sampled), fused by a cross-channel attention module.
Problem
Single-network implicit neural video representations (NeRV-style) entangle static and dynamic content, so they cannot efficiently compress redundant static information or model globally coherent dynamic details.
Approach
Decompose video into sparse learnable static codes sampled by distance-weighted sum of two neighboring codes, and dynamic codes obtained by interpolating a denser code set over time, without optical flow or residual supervision; fuse them with a cross-channel attention module before NeRV-style decoding.
Key results
- Reaches 31.20 PSNR on Bunny with 0.35M parameters, versus 29.78 for HNeRV at the same size.
- Achieves 33.31 average PSNR on UVG at 960x1920, versus 32.82 for HNeRV and 31.32 for NeRV.
- Achieves 30.36 average PSNR on DAVIS versus 28.91 for HNeRV; also best on inpainting and interpolation training-set results.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| decompose videos into learned static and dynamic codes | DS-NeRV · Big Buck Bunny, 0.35M parameters | PSNR: 29.78 -> 31.20 | Helped |
| decompose videos into learned static and dynamic codes | DS-NeRV · UVG 960×1920, average | PSNR: 32.82 -> 33.31 | Helped |
| decompose videos into learned static and dynamic codes | DS-NeRV · DAVIS, average | PSNR: 28.91 -> 30.36 | Helped |
| decompose videos into learned static and dynamic codes | DS-NeRV · DAVIS inpainting, central mask, average | PSNR: 23.03 -> 25.06 | Helped |
| increase dynamic-code length (ablation) | DS-NeRV · Jockey | PSNR: For static length 30: 28.68 -> 30.87 -> 32.16 -> 32.56 | Helped |
| increase dynamic-code length (ablation) | DS-NeRV · Honey | PSNR: For static length 30: 39.13 -> 39.53 -> 39.52 -> 39.43 | No clear effect |
| vary static-code length while dynamic-code length is 0 (ablation) | DS-NeRV · Jockey | PSNR: 29.91 -> 28.68 -> 29.91 -> 30.88 | Mixed |
| vary static-code length while dynamic-code length is 0 (ablation) | DS-NeRV · Honey | PSNR: 38.80 -> 39.13 -> 39.23 -> 39.38 | Helped |
| use both static and dynamic codes instead of setting one code length to 0 (ablation) | DS-NeRV · Jockey and Honey | PSNR | Helped |
Surprising findings
- Interpolation test frames are not seen in training, yet DS-NeRV still achieves comparable test-set PSNR, while HNeRV and DNeRV use the test frames themselves as encoder input.
Limitations
- Experiments limited to three video datasets (Bunny, UVG, DAVIS) at fixed resolutions.
- Requires separate training per video, as with other NeRV-style methods; static/dynamic code lengths must be tuned per video.
Experimental setup
Datasets: Big Buck Bunny UVG DAVIS
Benchmarks: Big Buck Bunny UVG DAVIS
Metrics: PSNR MS-SSIM bpp
Baselines: NeRV DNeRV HNeRV H.264 HEVC
Builds on: NeRV HNeRV DNeRV GLO E-NeRV
Closest papers by meaning
- Boosting Neural Representations for Videos with a Conditional Decoder (CVPR 2024 · similarity 0.90)
- Combining Frame and GOP Embeddings for Neural Video Representation (CVPR 2024 · similarity 0.89)
- Towards Practical Real-Time Neural Video Compression (CVPR 2025 · similarity 0.87)
- NVRC: Neural Video Representation Compression (NeurIPS 2024 · similarity 0.87)
- MetaNeRV: Meta Neural Representations for Videos with Spatial-Temporal Guidance (AAAI 2025 · similarity 0.87)
- CoANeRV: Coordinate-Aware Token-Space Neural Video Representation (NeurIPS 2026 · 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.