Research map / 3D vision / 3D generation
Neural Atlas Graphs for Dynamic Scene Decomposition and Editing
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 Neural Atlas Graphs, a hybrid scene graph where each object/background node is a view-dependent neural atlas on a 3D plane, enabling 2D appearance editing and 3D ordering/positioning, fit at test time on Waymo and DAVIS.
Problem
Editable high-resolution representations of dynamic scenes: neural atlases are editable in 2D but break down with multiple occluding interacting objects, while neural scene graphs handle 3D layout but are hard to edit view-consistently.
Approach
Represent each scene element as a 2D planar neural atlas with a learned rigid 3D trajectory, optional spline-based flow, and view-dependent color/alpha fields; render via depth-ordered ray-plane intersection and alpha compositing, using 3D bounding boxes and masks, so textures edited in atlas space propagate consistently.
Key results
- On 7 Waymo scene segments (25 subsequences), NAG reaches mean PSNR 41.85 vs 36.78 for OmniRe and 34.93 for EmerNeRF.
- On dynamic objects, NAG gains 11.2 dB PSNR for vehicles and 10.7 dB for humans over the best baseline.
- On 15 DAVIS sequences, NAG achieves mean PSNR 35.35 vs 28.09 for Layered Neural Atlases and 27.31 for OmnimatteRF (about 7 dB gain).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use NAG instead of OmniRe | Dynamic scene reconstruction · Waymo Open Dataset (mean) | PSNR: 36.78 -> 41.85 | Helped |
| Use NAG instead of OmniRe | Dynamic scene reconstruction · Waymo Open Dataset (mean) | SSIM: 0.967 -> 0.978 | Helped |
| Use NAG instead of OmniRe | Dynamic scene reconstruction · Waymo Open Dataset (mean) | LPIPS: 0.070 -> 0.051 | Helped |
| Use NAG instead of OmniRe | Vehicle rendering · Waymo vehicle objects (mean) | PSNR: 31.69 -> 42.88 | Helped |
| Use NAG instead of OmniRe | Vehicle rendering · Waymo vehicle objects (mean) | SSIM: 0.922 -> 0.986 | Helped |
| Use NAG instead of OmniRe | Human rendering · Waymo human objects (mean) | PSNR: 32.24 -> 42.94 | Helped |
| Use NAG instead of OmniRe | Human rendering · Waymo human objects (mean) | SSIM: 0.901 -> 0.981 | Helped |
| Use NAG instead of Layered Neural Atlases | Outdoor video reconstruction · DAVIS (mean) | PSNR: 28.09 -> 35.35 | Helped |
| Use NAG instead of Layered Neural Atlases | Outdoor video reconstruction · DAVIS (mean) | SSIM: 0.800 -> 0.929 | Helped |
| Use NAG instead of Layered Neural Atlases | Outdoor video reconstruction · DAVIS (mean) | LPIPS: 0.302 -> 0.098 | Helped |
Limitations
- Assumes objects and background can be projected onto planes, limiting camera and object rotation to under 180 degrees; novel view synthesis is out of scope.
- Initialization via bounding boxes or homographies is sensitive; suboptimal initialization can lead to early failure.
- Large camera motion that changes background content is difficult to represent with the planar assumption.
Experimental setup
Datasets: Waymo Open Dataset DAVIS
Benchmarks: Waymo Open Dataset DAVIS
Metrics: PSNR SSIM LPIPS
Baselines: OmniRe EmerNeRF Layered Neural Atlases OmnimatteRF
Builds on: Layered Neural Atlases Neural Scene Graphs Hermite spline trajectory Instant-NGP hash encoding
Closest papers by meaning
- EmerNeRF: Emergent Spatial-Temporal Scene Decomposition via Self-Supervision (ICLR 2024 · similarity 0.85)
- SceneCrafter: Controllable Multi-View Driving Scene Editing (CVPR 2025 · similarity 0.85)
- Learning to Evolve Scenes: Reasoning about Human Activities with Scene Graphs (NeurIPS 2026 · similarity 0.85)
- Multi-Level Neural Scene Graphs for Dynamic Urban Environments (CVPR 2024 · similarity 0.85)
- DORSal: Diffusion for Object-centric Representations of Scenes $\textit{et al.}$ (ICLR 2024 · similarity 0.85)
- A Closer Look at Dynamic Scene Graph Generation in the Era of Multimodal Large Language Models (NeurIPS 2026 · 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.