| ## Overall | |
| This is the dataset for [SeeU: Seeing the Unseen World via 4D Dynamics-aware Generation](https://yuyuanspace.com/SeeU/). | |
| ## Dataset Structure | |
| - `SeeU45_train/` | |
| Provides the 2D frames used as model inputs during training. | |
| - `SeeU45_GT/` | |
| Contains the full scene ground-truth frame sequences. | |
| - `SeeU45_GT/sample_frame.txt` | |
| Specifies which frames are sampled for training. It also records: | |
| - the original scene/video folder names, | |
| - sampled frame indices, | |
| - total frame counts (GT length), | |
| - and frame numbers corresponding to `SeeU45_train` inputs. | |
| ## Data Sources and Licensing | |
| SeeU45 is constructed from a combination of our own captured scenes and publicly available datasets. | |
| Due to licensing restrictions on some third-party datasets, we are not allowed to re-distribute certain videos. | |
| As a result, the public release of SeeU45 **omits 5 scenes** that appear in our paper. | |
| The publicly released scenes in SeeU45 are derived from the following sources (in compliance with their respective licenses): | |
| - [TAP-Vid](https://tapvid.github.io/#:~:text=The%20annotations%20of%20TAP-Vid%2C%20as%20well%20as%20the,their%20creators%3B%20see%20the%20DAVIS%20dataset%20for%20details.) | |
| - [AgiBot](https://huggingface.co/datasets/agibot-world/AgiBotWorld-Alpha) | |
| - [I2-2000FPS](https://chennuriprateek.github.io/Quanta_Video_Restoration-QUIVER-/) | |
| All original copyrights of these source datasets are retained by their respective authors. | |
| ## SeeU45 License | |
| The **SeeU45 dataset** (including our processed frames, splits, and annotations) is released under: | |
| **license: CC BY-NC-ND 4.0** | |
| This means: | |
| - ✅ You may use SeeU45 for **non-commercial academic research**. | |
| - ✅ You must credit the SeeU paper and dataset when using it. | |
| - ❌ You may not use SeeU45 for commercial purposes. |