Research map / 3D vision / NeRF and radiance fields
Multi-Level Neural Scene Graphs for Dynamic Urban Environments
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 a multi-level neural scene graph radiance field with composite ray sampling for scalable novel view synthesis in dynamic urban environments, plus a new Argoverse 2-based benchmark.
Problem
Novel view synthesis of large-scale dynamic urban areas from many vehicle captures under varying conditions, where prior methods are static, do not scale beyond a single short video, or fail to separate dynamic object instances.
Approach
A multi-level scene graph (root, sequence, object, camera nodes) with latent codes conditions two radiance fields: a static field on sequence latents and a dynamic-object field on object and sequence latents. Composite ray sampling samples the dynamic field only inside object bounding boxes, and hierarchical pose optimization refines ego-vehicle and object poses through graph edges.
Key results
- On the proposed Argoverse 2 benchmark, the method reaches 21.15 mean PSNR / 0.680 SSIM / 0.555 LPIPS versus SUDS at 20.84 / 0.662 / 0.601, while training in 17.2 h versus 54.8 h for SUDS.
- On KITTI with 75% training views, PSNR 28.38 and LPIPS 0.052, versus 23.12 PSNR for SUDS at 50% views (the best prior result shown in the table).
- Composite ray sampling trains about 12x faster than separate per-node sampling (30K vs 2.5K rays/sec) with comparable quality.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the multi-level neural scene graph and composite rendering | Dynamic urban radiance field · Argoverse 2 mean | PSNR: 20.84 -> 21.15 | Helped |
| Use the multi-level neural scene graph and composite rendering | Dynamic urban radiance field · Argoverse 2 mean | LPIPS: 0.601 -> 0.555 | Helped |
| Use the proposed method instead of the strongest baseline | Dynamic urban radiance field · KITTI [75%] | PSNR: 21.53 -> 28.38 | Helped |
| Use the proposed method instead of the strongest baseline | Dynamic urban radiance field · VKITTI2 [75%] | PSNR: 23.87 -> 29.73 | Helped |
| Add sequence-level appearance embeddings (ablation) | Multi-level scene graph radiance field · Argoverse 2 residential, single sequence | PSNR: 19.70 -> 22.22 | Helped |
| Add transient-geometry embeddings to appearance embeddings (ablation) | Multi-level scene graph radiance field · Argoverse 2 residential, single sequence | PSNR: 22.22 -> 22.49 | Helped |
| Replace composite ray sampling with uniform sampling (ablation) | Dynamic urban radiance field · Argoverse 2 residential, single sequence | PSNR: 27.07 -> 25.72 | Hurt |
| Replace composite ray sampling with separate per-node sampling (ablation) | Dynamic urban radiance field · Argoverse 2 residential, single sequence | rays/sec: 30K -> 2.5K | Hurt |
| Use hierarchical instead of naive pose optimization (ablation) | Dynamic urban radiance field · Argoverse 2 residential | PSNR: 21.28 -> 22.29 | Helped |
| Use hierarchical instead of naive pose optimization (ablation) | Dynamic urban radiance field · Argoverse 2 residential | SSIM: 0.663 -> 0.678 | Helped |
Surprising findings
- Naive camera pose optimization substantially degrades PSNR (22.49 to 21.28), while hierarchical pose optimization keeps PSNR comparable and improves SSIM and LPIPS.
Limitations
- Relies on provided 3D bounding box tracks; an off-the-shelf tracker is only evaluated in supplementary material.
- Ablations on graph structure and sampling are run on a single sequence of the residential area because full-scale variants are prohibitively expensive.
Experimental setup
Datasets: Argoverse 2 KITTI VKITTI2
Benchmarks: Argoverse 2 NVS benchmark KITTI VKITTI2
Metrics: PSNR SSIM LPIPS training time
Baselines: Nerfacto + Emb. Nerfacto + Emb. + Time SUDS NSG NeRF NeRF + Time SRN PNF
Builds on: Mip-NeRF 360 / proposal sampling (Nerfacto) Neural Scene Graphs (NSG) Instant-NGP hash grid NeRF SUDS
Closest papers by meaning
- Dynamic 3D Gaussian Fields for Urban Areas (NeurIPS 2024 · similarity 0.91)
- BLiRF: Bandlimited Radiance Fields for Dynamic Scene Modeling (AAAI 2024 · similarity 0.88)
- DetRF: Detachable Novel Views Synthesis of Dynamic Scenes Using Backdrop-Driven Neural Radiance Fields (AAAI 2025 · similarity 0.87)
- LidaRF: Delving into Lidar for Neural Radiance Field on Street Scenes (CVPR 2024 · similarity 0.87)
- NeuRAD: Neural Rendering for Autonomous Driving (CVPR 2024 · similarity 0.86)
- DynamicCity: Large-Scale 4D Occupancy Generation from Dynamic Scenes (ICLR 2025 · 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.