Oral Presentation, 3DV 2021

Video: Youtube

We present Hand ArticuLated Occupancy (HALO), a novel representation of articulated hands that bridges the advantages of 3D keypoints and neural implicit surfaces and can be used in end-to-end trainable architectures. Unlike existing statistical parametric hand models (e.g.~MANO), HALO directly leverages the 3D joint skeleton as input and produces a neural occupancy volume representing the posed hand surface. The key benefits of HALO are (1) it is driven by 3D keypoints, which have benefits in terms of accuracy and are easier to learn for neural networks than the latent hand-model parameters; (2) it provides a differentiable volumetric occupancy representation of the posed hand; (3) it can be trained end-to-end, allowing the formulation of losses on the hand surface that benefit the learning of 3D keypoints. We demonstrate the applicability of HALO to the task of conditional generation of hands that grasp 3D objects. The differentiable nature of HALO is shown to improve the quality of the synthesized hands both in terms of physical plausibility and user preference.

• December 1, 2021: Initial release for version 0.01 with demo.

• im2mesh library from Occupancy Network for point sampling and marching cubes
• requirements.txt
• MANO model for data preprocessing

The easiest way to run the code is to use conda. The code is tested on Ubuntu 18.04.

To try a demo which produces an implicit hand surface from the input keypoints, run:

cd halo
python demo_kps_to_hand.py

The demo will run the marching cubes algorithm and render each image in the animation above sequentially. The output images are in the output folder. The provided sample sequence are interpolations beetween 17 randomly sampled poses from the unseen HO3D dataset .

• The HALO-base model is trained using Youtube3D hand dataset. We only use the hand mesh ground truth without the images and videos. We provide the preprocessed data in the evaluation section.
• The HALO-VAE model is trained and test on the GRAB dataset

To generate the mesh given the 3D keypoints and precomputed transformation matrices, run:

cd halo_base
python generate.py CONFIG_FILE.yaml

To evaluate the hand surface, run:

python eval_meshes.py

We provide the preprocessed test set of the Youtube3D here. In addition, you can also find the produced meshes from our keypoint model on the same test set here.

To generate grasps given 3D object mesh, run:

python generate.py HALO_VAE_CONFIG_FILE.ymal --test_data DATA_PATH --inference

The evaluation code for contact/interpenetration and cluster analysis can be found in halo/evaluate.py and halo/evaluate_cluster.py accordningly. The intersection test demo is in halo/utils/interscetion.py

Each data point consists of 3D keypoints, transformation matrices, and a hand surface. To speed up the training, all transformation matrices are precomputed, either by out Canonicalization Layer or from the MANO. Please check halo/halo_base/prepare_data_from_mano_param_keypoints.py for details. After the metadata is processed, sample points and compute occupancy values with bash build_dataset.sh. We use the surface point sampling and occupancy computation method from the Occupancy Networks

To train HALO base model (implicit functions), run:

cd halo_base
python train.py

To train HALO-VAE, run:

cd halo
python train.py

HALO_VAE requires a HALO base model trained using the transformation matrices from the Canonicalization Layer. The weights of the base model are not updated during the VAE training.

@inproceedings{karunratanakul2021halo,
  title={A Skeleton-Driven Neural Occupancy Representation for Articulated Hands},
  author={Karunratanakul, Korrawe and Spurr, Adrian and Fan, Zicong and Hilliges, Otmar and Tang, Siyu},
  booktitle={International Conference on 3D Vision (3DV)},
  year={2021}
}

Some code in our repo uses snippets of the following repo:

• The occupancy computation and point sampling is based on Occupancy Networks.
• We use MANO layer implementation from Yana Hasson
• We use the renderer from MeshTransformer in our demo.

Please consider citing them if you found the code useful.

We sincerely acknowledge Shaofei Wang and Marko Mihajlovic for the insightful discussionsand helps with the baselines.