chkim403/gs_physics

GS Physics Dataset

This dataset accompanies the CVPR 2026 Findings paper Learning a Particle Dynamics Model with Real-world Videos. It contains real-world multi-object interaction recordings and processed Gaussian Splatting particle-dynamics data for the bowling and cube_stacks scenarios.

The release includes two archive files:

• dataset.zip: raw multi-view RGB/stereo/depth recordings, segmentation masks, camera metadata, and object-pose estimates.
• processed_dataset.zip: processed data used by the released training and evaluation code, including masked RGB images, segmentation masks, object poses, and per-frame Gaussian parameter files.

The official code repository is available at: chkim403/gs-particle-dynamics.

Dataset Summary

The dataset contains 529 real-world scenes captured from four synchronized camera views:

Scenario	Number of scenes
bowling	297
cube_stacks	232
Total	529

For the experiments in the paper, 292 bowling scenes and 210 cube_stacks scenes were used. The remaining scenes were excluded because of issues observed in their preprocessed outputs, such as segmentation failures. The excluded-scene list is available as skip_list in tools/dataset.py in the official code repository.

Scene IDs are not guaranteed to be continuous because some unusable scenes were manually removed after continuous recording.

Files

.
├── dataset.zip
└── processed_dataset.zip

Both archives include the dataset license and terms of use.

Raw Dataset

dataset.zip contains the original released dataset files.

Raw root structure

.
├── bowling/
│   ├── scene_00001/
│   ├── scene_00002/
│   └── ...
├── cube_stacks/
│   ├── scene_00003/
│   ├── scene_00004/
│   └── ...
├── dataset_license.md
├── dataset_terms_of_use.md
└── raw_dataset.md

Each scene contains four camera folders:

<scenario>/scene_XXXXX/
├── 234322305266/
├── 248622303451/
├── 336222300744/
└── 336522303601/

The same four camera IDs are used across all released scenes.

Raw camera directory structure

Each camera folder contains RGB images, stereo images, stereo depth aligned to RGB pixels, segmentation masks, object poses, and per-camera metadata:

<scenario>/scene_XXXXX/<camera_id>/
├── camera_meta.json
├── obj_poses.npz
├── rgb/
├── left/
├── right/
├── stereo_aligned_depth/
└── filtered_segmentation_DAM4SAM/

camera_meta.json is a per-camera JSON file:

{
  "h": 480,
  "w": 640,
  "k": [[...], [...], [...]],
  "w2c": [[...], [...], [...], [...]],
  "depth_scale": 0.0010000000474974513
}

Fields:

• h, w: image height and width.
• k: 3x3 RGB camera intrinsic matrix.
• w2c: 4x4 world-to-camera extrinsic matrix.
• depth_scale: multiplier for converting stored depth values to metric depth, generated from the Intel RealSense API.

Additional raw files:

• obj_poses.npz: pseudo object pose information for the corresponding camera.
• rgb/, left/, right/: zero-padded .jpg image sequences.
• stereo_aligned_depth/: zero-padded .npy depth arrays aligned to the RGB camera.
• filtered_segmentation_DAM4SAM/: zero-padded .png segmentation masks.

RGB, stereo, and depth streams contain full camera sequences. Segmentation masks cover selected interaction ranges chosen by human annotators and may contain fewer frames than the full 120-frame RGB/depth streams.

Segmentation, depth, and object poses were generated by external algorithms described in the paper and should be treated as pseudo annotations rather than perfect ground truth. Object IDs in segmentation masks are intended to be consistent across time and views, but masks may contain noise from tracking and cross-view association.

Processed Dataset

processed_dataset.zip contains the data required by the released training and evaluation code.

Processed root structure

.
├── bowling/
│   ├── train/
│   └── test/
├── cube_stacks/
│   ├── train/
│   └── test/
├── dataset_license.md
├── dataset_terms_of_use.md
└── processed_dataset.md

Processed scene structure

Each processed scene has scene-level camera/image metadata plus a gs/ folder containing per-frame Gaussian Splatting parameter files:

<scenario>/<split>/scene_XXXXX/
├── camera_meta.json
├── rgb/
├── seg/
├── obj_poses/
└── gs/
    ├── <frame>/
    │   ├── params_coarse.npz
    │   └── gs_soft_ids_coarse.npz
    └── ...

<split> is either train or test.

camera_meta.json contains camera metadata shared by the scene:

{
  "h": 480,
  "w": 640,
  "cam_id": [
    "234322305266",
    "248622303451",
    "336222300744",
    "336522303601"
  ],
  "k": [...],
  "w2c": [...]
}

Processed scene files:

• rgb/<camera_id>/*.png: masked RGB images that keep only foreground objects.
• seg/<camera_id>/*.png: segmentation masks used to produce the masked RGB images.
• obj_poses/<camera_id>/obj_poses.npz: per-camera object pose data.
• gs/<frame>/params_coarse.npz: Gaussian parameters for that frame.
• gs/<frame>/gs_soft_ids_coarse.npz: object-ID assignments for the Gaussians.

The rgb/ and seg/ folders contain the selected interaction frame range. They contain three more frames than the corresponding gs/ folder, with the extra frames appended at the end of the sequence. Gaussian parameters are not stored for these final frames because the input Gaussian trajectories used by the model are generated by applying the transformations stored in obj_poses/ to static Gaussians.

License

Unless otherwise stated, the dataset is licensed under the Creative Commons Attribution-NonCommercial 4.0 International License (CC BY-NC 4.0). Code, scripts, and software in the official repository are licensed separately under the MIT License.

By downloading, accessing, or using the dataset, you are responsible for complying with the dataset license, the dataset terms of use, and all applicable laws and regulations.

Citation

If you use this dataset in a publication, project, benchmark, or released model, please cite:

@inproceedings{kim2026learning,
  title     = {Learning a Particle Dynamics Model with Real-world Videos},
  author    = {Kim, Chanho and Sumukh, Suhas V. and Fuxin, Li},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Findings},
  year      = {2026}
}

Contact

For questions, please contact:

Chanho Kim
kimchanh@oregonstate.edu