GraspTTA

Hand-Object Contact Consistency Reasoning for Human Grasps Generation (ICCV 2021).

Demo

Quick Results Visualization

We provide generated grasps on out-of-domain HO-3D dataset (saved at ./diverse_grasp/ho3d), you can visualize the results by:

python vis_diverse_grasp --obj_id=6

The visualization will look like this:

Generate diverse grasps on out-of-domain HO-3D dataset (the model is trained on ObMan dataset)

You can also generate the grasps by yourself

First, download pretrained weights, unzip and put into checkpoints.
Second, download the MANO model files (mano_v1_2.zip) from MANO website. Unzip and put mano/models/MANO_RIGHT.pkl into models/mano.
Third, download HO-3D object models, unzip and put into models/HO3D_Object_models.
The structure should look like this:

GraspTTA/
  checkpoints/
    model_affordance_best_full.pth
    model_cmap_best.pth
  models/
    HO3D_Object_models/
      003_cracker_box/
        points.xyz
        textured_simple.obj
        resampled.npy
       ......
    mano/
      MANO_RIGHT.pkl

Then, install the V-HACD for building the simulation of grasp displacement. Change this line to your own path.
Finally, run run.sh for installing other dependencies and start generating grasps.

Generate grasps on custom objects

First, resample 3000 points on object surface as the input of the network. You can use this function.
Second, write your own dataloader and related code in gen_diverse_grasp_ho3d.py.

Training code

Upsate soon

Citation

@inproceedings{jiang2021graspTTA,
          title={Hand-Object Contact Consistency Reasoning for Human Grasps Generation},
          author={Jiang, Hanwen and Liu, Shaowei and Wang, Jiashun and Wang, Xiaolong},
          booktitle={Proceedings of the International Conference on Computer Vision},
          year={2021}
}

Acknowledgments

We thank:

MANO provided by Omid Taheri.
This implementation of PointNet.
This implementation of CVAE.

Umesato/GraspTTA