The ARCTIC dataset and repository provide tools for downloading, processing, visualizing, and training models on dexterous bimanual hand-object manipulation. It targets researchers and developers in computer vision and robotics, offering a large-scale dataset with rich annotations for advancing hand-object interaction research.

How It Works

The project offers a comprehensive pipeline for the ARCTIC dataset, including scripts for data preprocessing, splitting, and rendering (RGB, depth, segmentation masks). It also provides baseline models and a generalized codebase for training, visualization, and evaluation of custom models, facilitating reproduction of research and development of new methods in hand-object interaction.

Quick Start & Requirements

• Install via git clone https://github.com/zc-alexfan/arctic.git.
• Detailed setup instructions are available in docs/setup.md.
• Data download and visualization instructions are in docs/data/README.md.
• Training and evaluation instructions for baselines are in docs/model/README.md.

Highlighted Details

• Contains 2.1M high-resolution images with annotated frames.
• Features 3D groundtruth for SMPL-X, MANO, and articulated objects.
• Captured using 54 high-end Vicon cameras in a MoCap setup.
• Supports tasks like template-free bimanual hand-object reconstruction and egocentric hand-object reconstruction.

Maintenance & Community

• The project is associated with CVPR 2023 and has an ECCV'24 competition.
• For technical questions, create an issue. For other inquiries, contact arctic@tue.mpg.de.

Licensing & Compatibility

• The dataset and code are released under a license detailed in LICENSE.
• For commercial licensing, contact ps-licensing@tue.mpg.de.

Limitations & Caveats

The README mentions a CVPR 2024 Highlight for HOLD, which jointly reconstructs hands and objects without assuming a pre-scanned object template or 3D hand-object training data, suggesting potential advancements beyond the initial ARCTIC dataset's scope.