Research map / 3D vision / 3D generation
DiVa-360: The Dynamic Visual Dataset for Immersive Neural Fields
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.
Introduces DiVa-360, a real-world 360-degree multi-view dynamic dataset of 53-camera table-scale scenes (17.4M frames), and benchmarks per-frame I-NGP, MixVoxels, and K-Planes on it.
Problem
Dynamic neural fields lag behind 2D video for long-duration, high-fidelity capture, partly due to the lack of large-scale real-world multi-view datasets with long, 360-degree coverage.
Approach
A low-cost 53-camera capture rig (BRICS) produces synchronized 120 FPS, 360-degree, long-duration table-scale sequences with multi-view-consistent I-NGP-derived foreground masks, and standardized benchmarks show per-frame I-NGP can beat dedicated dynamic NeRF methods.
Key results
- PF I-NGP achieves 28.31 PSNR versus 27.68 for MixVoxels and 26.39 for K-Planes on DiVa-360, despite not using temporal information.
- I-NGP-based foreground segmentation reaches 0.926 mIoU versus 0.919 for SAM on DiVa-360.
- MixVoxels PSNR drops from 27.75 to 27.2 and 26.91 to 26.1 with added temporal length, while K-Planes improves from 25.57 to 26.04 and 26.03 to 26.19.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train each frame separately without adjacent-frame temporal information | PF I-NGP versus MixVoxels and K-Planes · DiVa-360 | PSNR: 27.68 ± 2.51 -> 28.31 ± 3.27; 26.39 ± 3.13 -> 28.31 ± 3.27 | Helped |
| Train each frame separately without adjacent-frame temporal information | PF I-NGP versus MixVoxels and K-Planes · DiVa-360 | training time (s/frame): 57.55 ± 6.96 -> 48.70 ± 4.40; 47.59 ± 5.13 -> 48.70 ± 4.40 | Mixed |
| Use fewer chunks (more temporal information) (ablation) | K-Planes · DiVa-360 object and interaction sequences | PSNR | Helped |
| Use fewer chunks (more temporal information) (ablation) | MixVoxels · DiVa-360 object and interaction sequences | PSNR | Hurt |
| Provide more temporal information using 2 or 3 chunks (ablation) | K-Planes versus MixVoxels · DiVa-360 interaction data | PSNR | Helped |
| Train at lower image resolution (ablation) | PF I-NGP · DiVa-360 | PSNR: 28.16 -> 28.19 -> 28.15 | No clear effect |
| Train at lower image resolution (ablation) | MixVoxels · DiVa-360 | PSNR: 27.63 -> 27.75 -> 26.91 | Mixed |
| Train at lower image resolution (ablation) | K-Planes · DiVa-360 | PSNR: 25.38 -> 25.57 -> 26.03 | Helped |
| Use lower spatial resolution and longer temporal length (464 × 220, 1 chunk) (ablation) | K-Planes versus 1160 × 550, 6 chunks · DiVa-360 | PSNR: 25.38 -> 26.19 | Helped |
| Use lower spatial resolution and longer temporal length (464 × 220, 1 chunk) (ablation) | MixVoxels versus 1160 × 550, 6 chunks · DiVa-360 | PSNR: 27.63 -> 26.10 | Hurt |
| Use all-view 360° cameras instead of forward-facing cameras (ablation) | PF I-NGP · DiVa-360 occluded views | PSNR: 23.99 -> 27.24 | Helped |
| Use all-view 360° cameras instead of forward-facing cameras (ablation) | MixVoxels · DiVa-360 occluded views | PSNR: 15.80 -> 27.43 | Helped |
Surprising findings
- Per-frame PF I-NGP, which ignores temporal information, outperforms dedicated dynamic methods MixVoxels and K-Planes in reconstruction quality and training speed.
- Lower spatial resolution training leaves PF I-NGP quality almost unchanged and improves MixVoxels and K-Planes, contradicting the expectation that higher resolution helps.
Limitations
- BRICS only captures table-scale volumes (about 1 m^3), not larger scenes.
- Only three dynamic neural field baselines were benchmarked due to slow training of state-of-the-art methods on long sequences.
- Current metrics and evaluation are limited to images; audio and text are not evaluated.
Experimental setup
Datasets: DiVa-360
Benchmarks: DiVa-360
Metrics: PSNR SSIM LPIPS JOD train time per frame render time per frame mIoU
Baselines: PF I-NGP MixVoxels K-Planes SAM
Builds on: Instant-NGP MixVoxels K-Planes COLMAP
Closest papers by meaning
- From an Image to a Scene: Learning to Imagine the World from a Million 360° Videos (NeurIPS 2024 · similarity 0.86)
- 4D3R: Motion-Aware Neural Reconstruction and Rendering of Dynamic Scenes from Monocular Videos (NeurIPS 2025 · similarity 0.86)
- ResFields: Residual Neural Fields for Spatiotemporal Signals (ICLR 2024 · similarity 0.85)
- MotionBench: Benchmarking and Improving Fine-grained Video Motion Understanding for Vision Language Models (CVPR 2025 · similarity 0.85)
- DL3DV-10K: A Large-Scale Scene Dataset for Deep Learning-based 3D Vision (CVPR 2024 · similarity 0.85)
- Real-time Photorealistic Dynamic Scene Representation and Rendering with 4D Gaussian Splatting (ICLR 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.