Datasets:
| license: cc-by-4.0 | |
| task_categories: | |
| - image-to-video | |
| tags: | |
| - simulation | |
| - clevr | |
| - objects | |
| - 3d | |
| - video | |
| - prediction | |
| # OBJ3D Dataset | |
| OBJ3D dataset originally from [G-SWM](https://github.com/zhixuan-lin/G-SWM). It contains video frames of synthetic CLEVR-like objects colliding. | |
| This dataset is used for evaluating object-centric world models and stochastic dynamics, as featured in the paper: **[Latent Particle World Models: Self-supervised Object-centric Stochastic Dynamics Modeling](https://huggingface.co/papers/2603.04553)**. | |
| - **Project Page:** [https://taldatech.github.io/lpwm-web](https://taldatech.github.io/lpwm-web) | |
| - **GitHub Repository:** [https://github.com/taldatech/lpwm](https://github.com/taldatech/lpwm) | |
| ## Sample Usage | |
| To train models on this dataset using the official implementation, you can use the following commands: | |
| ### Single-GPU Training (DLPv3) | |
| ```bash | |
| python train_dlp.py --dataset obj3d_img | |
| ``` | |
| ### Multi-GPU Training (LPWM) | |
| ```bash | |
| accelerate launch --config_file ./accel_conf.yml train_lpwm_accelerate.py --dataset obj3d128 | |
| ``` | |
| ## Citation | |
| ```bibtex | |
| @inproceedings{ | |
| daniel2026latent, | |
| title={Latent Particle World Models: Self-supervised Object-centric Stochastic Dynamics Modeling}, | |
| author={Tal Daniel and Carl Qi and Dan Haramati and Amir Zadeh and Chuan Li and Aviv Tamar and Deepak Pathak and David Held}, | |
| booktitle={The Fourteenth International Conference on Learning Representations}, | |
| year={2026}, | |
| url={https://openreview.net/forum?id=lTaPtGiUUc} | |
| } | |
| ``` |