ArGS

KO | EN

gitlite — search

#javascript #python #go #android #deep-learning #docker #nodejs #windows #awesome #kubernetes #vue #pytorch #security #mcp

(★ 21)

[CVPR 2026 Highlight] Clay-to-Stone: Phase-wise 3D Gaussian Splatting for Monocular Articulated Hand-Object Manipulation Modeling

ArGS 최신버젼 다운로드

최종 버전 다운로드 (.zip)

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

<h1>Clay-to-Stone: Phase-wise 3D Gaussian Splatting for Monocular Articulated Hand-Object Manipulation Modeling [CVPR 2026 Highlight]</h1>

</div>

<table align="center" style="width:100%;">
  <tr>
    <td align="center" style="width:50%;">
      <img src="assets/yearbook.gif" width="100%"/>
      <br>
      <em>In-the-wild Video</em>
    </td>
    <td align="center" style="width:50%;">
      <img src="assets/render.gif" width="100%"/>
      <br>
      <em>Renders</em>
    </td>
  </tr>
</table>

<p align="center">
  <a href="https://youtu.be/1SULG-b25Uk"><b>🎥 Video Demo</b></a> |
  <a href="https://openaccess.thecvf.com/content/CVPR2026/papers/Liu_Clay-to-Stone_Phase-wise_3D_Gaussian_Splatting_for_Monocular_Articulated_Hand-Object_Manipulation_CVPR_2026_paper.pdf"><b>📄 Paper</b></a>
</p>

## Requirements

Refer to [HOGS](https://github.com/ru1ven/HOGS/blob/main/requirements.txt) for CUDA, PyTorch, and other dependencies. Additionally, install [gaussian-splatting](https://github.com/graphdeco-inria/gaussian-splatting.git).

## Dataset Preparation

Download the [ARCTIC](https://github.com/zc-alexfan/arctic) dataset and preprocess it using `cocoify_arctic.py` and `preprocess_arctic.py` to generate the training and testing splits used in the paper.

Generate part-level object segmentation for articulated sequences using [SAM2](https://github.com/facebookresearch/sam2).

[Here](https://huggingface.co/ru1ven/ArGS/tree/main) are our pre-trained checkpoints and segmented masks on ARCTIC.
<p align="center">
  <img src="assets/teaser_sam2.png" width="450px"/>
</p>

After preprocessing, the dataset should be organized as follows:

```text
arctic/
├── images/
├── masks/
└── splits/
    ├── train/
    │   └── *.npy
    └── test/
        └── *.npy
```

## Training
Run the training script:

```bash
python train_arctic.py 
```
Training configurations can be adjusted in `./configs`.

## In-the-Wild Demo

The [`Ablation branch`](https://github.com/ru1ven/ArGS/tree/Ablation) contains the code for our ablation studies and the in-the-wild demo. For a monocular articulated manipulation video, the pipeline is as follows:

1. Use [SAM2](https://github.com/facebookresearch/sam2) to extract part-level segmentation of the articulated object.
2. Use [HAMER](https://github.com/geopavlakos/hamer) and [FoundationPose](https://github.com/NVlabs/FoundationPose) to predict hand-object poses. Since FoundationPose can only track rigid objects, static parts (e.g., the base of a stapler) from SAM2 segmentation can be provided as input to extract the corresponding pose. Create the corresponding `splits/*.npy` files following Dataset Preparation.
3. (Optional but recommended) Roughly estimate and initialize the articulated object’s [axis](https://github.com/ru1ven/ArGS/blob/945367421c578d92675cc472f4e16a9aff586f4a/models/deformer/rigid.py#L529). This trick is intentionally omitted in our experiments to keep the evaluation free of such heuristics, but it can improve robustness in practical in-the-wild scenarios.
4. Use `train_wild.py` and `render_wild.py` from the [`Ablation branch`](https://github.com/ru1ven/ArGS/tree/Ablation) to model the articulated objects and render results.

## Acknowledgement
We thank [3DGS](https://github.com/graphdeco-inria/gaussian-splatting), [3DGS-Avatar](https://github.com/mikeqzy/3dgs-avatar-release), and [ARCTIC](https://github.com/zc-alexfan/arctic) for their great works!

ArGS

ArGS 최신버젼 다운로드

오늘의 추천 프로젝트

Was this content helpful?

Recent Feedback

ArGS

ArGS 최신버젼 다운로드

오늘의 추천 프로젝트

Was this content helpful?

Recent Feedback

Download README