Lixin Yang* · Kailin Li* · Xinyu Zhan* · Fei Wu · Anran Xu . Liu Liu · Cewu Lu

This repo contains the OakInk data toolkit (oikit) -- a Python package that provides data loading, splitting, and visualization tools for the OakInk knowledge repository.

OakInk contains three parts:

• OakBase: Object Affordance Knowledge (Oak) base, including objects' part-level segmentation and attributes.
• OakInk-Image: a video dataset with 3D hand-object pose and shape annotations.
• OakInk-Shape: a 3D grasping pose dataset with hand and object mesh models.

• It contains 3D models, part segmentation, and affordance labels of 1,800 common household objects.
• It records human grasps with 100 (from 1,800) objects based on their affordances.
  • It contains 792 multi-view video clips (230K images) complemented with annotation.
  • Images are from four third-person views.
  • It contains dynamic grasping and handover motions.
  • It includes 3D ground-truth for MANO and objects.
• It contains a total of 50k hand-object interaction pose pairs involving the 1,800 objects.
  • 1k are from the recording, 49K are done via interaction transfer.

• For studying hand-object pose estimation and hand-held object reconstruction.
• For generating grasping pose, motion or handover with objects.
• For generating affordance-aware pose or motion for object manipulation.
• For transferring hand pose or motion to a new object.

Clone the repo

$ git clone https://github.com/lixiny/OakInk.git

• Install environment: see docs/install.md
• Get datasets: see docs/datasets.md

# visualize OakInk-Image mesh on sequence level:
#   * --draw_mode [mesh, wireframe] to switch between mesh and wireframe
#   * --seq_id: select sequence id from OAKINK_DIR/image/anno/seq_status.json to visualize
#   * --view_id: select from [0, 1, 2, 3] for visualize from different views.
python scripts/viz_oakink_image_seq.py --draw_mode mesh --view_id 1

# use OakInkImage to load data_split: train, mode: subject (SP1) and visualize:
#   * --data_split: select from [train, val, test, all]
#   * --mode_split: select from [default, object, subject, handobject]
python scripts/viz_oakink_image.py --data_split train --mode_split subject

# use OakInkShape to load object category: teapot and intent: use:
#   * --categories: select from OAKINK_DIR/shape/metaV2/yodaobject_cat.json, or "all"
#   * --intent_mode: select from [use, hold, liftup, handover] or "all"
#   * --data_split: select from [train, val, test, all]
python scripts/viz_oakink_shape.py --categories teapot --intent_mode use
# press `N` to load next sample

# use OakInkShape to load all the training grasps
python scripts/viz_oakink_shape.py --categories all --data_split train

# use OakInkShape to load all the training grasps in handover
python scripts/viz_oakink_shape.py --categories all --data_split train --intent_mode handover

• Hand Mesh Recovery: recovery hand mesh from monocular image.
• Grasp Generation: generate human-like hand mesh grasping a given object.

If you find OakInk dataset and oikit useful for your research, please considering cite us:

@inproceedings{YangCVPR2022OakInk,
  author    = {Yang, Lixin and Li, Kailin and Zhan, Xinyu and Wu, Fei and Xu, Anran and Liu, Liu and Lu, Cewu},
  title     = {{OakInk}: A Large-Scale Knowledge Repository for Understanding Hand-Object Interaction},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year      = {2022},
}