File size: 3,194 Bytes
2e6f267 53ed7ea 2e6f267 53ed7ea 2e6f267 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 | ---
task_categories:
- other
---
# 4DCodeBench-RealWorld
This dataset is part of [4DCodeBench](https://4dcodebench.com), a benchmark for evaluating coding agents on inverse graphics of dynamic scenes. It contains 100 real-world cases, each with a video of a physical event and a hand-annotated mask of the pixels that move in it. 23 of the videos cannot be redistributed; one script rebuilds them from their public sources.
**Resources:**
- Paper: [4DCodeBench: Benchmarking Agents on Inverse Graphics of Dynamic Scenes](https://huggingface.co/papers/2610.03715)
- Project page: [https://4dcodebench.com](https://4dcodebench.com)
- Code: [https://github.com/4DCodeBench/4DCodeBench](https://github.com/4DCodeBench/4DCodeBench)
## Layout
```text
videos/<case>.mp4 the reference video, as the benchmark reads it
annotations/<case>/dynamic_mask.npz the dynamic mask
metadata.jsonl one row per case
scripts/ prepare_videos.py and what it needs
```
135 MB of video and 95 MB of masks. Every video has at most 300 frames, at most 30 fps and a long side of at most 1280 px. `dynamic_mask.npz` holds `mask`, uint8 (frames, height, width) with 1 on a moving pixel, at the video's frame count and resolution.
## Completing the videos
```bash
pip install numpy opencv-python # plus ffmpeg and ffprobe on PATH
python scripts/prepare_videos.py
```
The script downloads the 23 sources into `downloads/` and writes `videos/<case>.mp4` with the same steps that made the benchmark. Frame count, frame rate and resolution match exactly; pixels can differ slightly with the H.264 encoder. The benchmark's `download_data.py` runs this step itself.
## Metadata
Each row of `metadata.jsonl` has `case`, `dataset`, `availability` (`included`, or `script` for the 23 rebuilt videos), `source`, `width`, `height`, `fps`, `frames`, `categories` (the materials involved) and `description` (the physical event). `source` is the URL of a web video and the file path inside its dataset otherwise; YouTube rows also carry `channel`, `title` and `license`.
## Sources
| dataset | videos | link |
|---|---|---|
| WISA-80K | 33 | https://github.com/360CVGroup/WISA |
| Web (Pexels, Mixkit, Pixabay, YouTube) | 28 | per video in `metadata.jsonl` |
| Physics-IQ | 12 | https://github.com/google-deepmind/physics-IQ-benchmark |
| ABC-130K | 5 | https://huggingface.co/datasets/XDOF/ABC-130k |
| T-REX | 4 | https://huggingface.co/datasets/zekaiwang/trex_dataset |
| Phys101 | 3 | http://phys101.csail.mit.edu/ |
| Phys-AD | 3 | https://guyao2023.github.io/Phys-AD/ |
| Robo360 | 3 | https://arxiv.org/abs/2312.06686 |
| AgiBot World | 2 | https://huggingface.co/datasets/agibot-world/AgiBotWorld-Alpha |
| RoboCook | 2 | https://hshi74.github.io/robocook/ |
| RoboCraft | 2 | http://hxu.rocks/robocraft/ |
| ALOHA Unleashed | 1 | https://aloha-unleashed.github.io/ |
| ManipArena | 1 | https://github.com/maniparena/maniparena-repo |
| TaskLevel-ILC | 1 | https://www.youtube.com/watch?v=FLiILOyQQbw |
The five web videos from YouTube are used under CC BY, from the channels Morten Møller, Philip Yecko, ASMR City and Madly Satisfying. |