Yana Hasson, Bugra Tekin, Federica Bogo, Ivan Laptev, Marc Pollefeys, and Cordelia Schmid

• Project page
• Paper

• Setup
  • Download and install code
  • Download datasets
• Demo
• Training
  • Train model for joint hand-object pose estimation
  • Train in sparsely annotated setting on FPHAB dataset
• Acknowledgements

• Retrieve the code

git clone https://github.com/hassony2/handobjectconsist`
cd handobjectconsist

• Create and activate the virtual environment with python dependencies

conda env create --file=environment.yml
conda activate handobject_env

• Go to MANO website

• Create an account by clicking Sign Up and provide your information

• Download Models and Code (the downloaded file should have the format mano_v*_*.zip). Note that all code and data from this download falls under the MANO license.

• unzip and copy the content of the models folder into the assets/mano folder

• Your structure should look like this:

handobjectconsist/
  assets/
    mano/
      MANO_LEFT.pkl
      MANO_RIGHT.pkl
      fhb_skel_centeridx9.pkl

• Download the First-Person Hand Action Benchmark dataset following the official instructions to the data/fhbhands folder
• Unzip the Object_models

unzip data/fhbhands/Object_models.zip -d data/fhbhands

• Unzip MANO fits

tar -xvf assets/fhbhands_fits.tgz -C assets/

• Download pre-trained models

wget https://github.com/hassony2/handobjectconsist/releases/download/v0.2/releasemodels.zip

unzip releasemodels.zip

• Optionally, resize the images (speeds up training !)

  • python reduce_fphab.py

• Your structure should look like this:

data/
  fhbhands/
    Video_files/
    Video_files_480/  # Optional, created by reduce_fphab.py script
    Subects_info/
    Object_models/
    Hand_pose_annotation_v1/
    Object_6D_pose_annotation_v1_1/
assets/
  fhbhands_fits/
releasemodels/
  fphab/
     ...

Note that all results in our paper are reported on a subset of the current dataset which was published as an early release, additionally we used synthetic data which is not released. The results are therefore not directly comparable with the final published results which are reported on the v2 version of the dataset.

After submisison I retrained a baseline model on the current dataset (official release of HO3D, which I refer to as HO3D-v2). You can get the model from the releasemodels

Evaluate the pre-trained model:

• Download pre-trained models

• Extract the pre-trained models unzip releasemodels.zip

• Run the evaluation code and generate the codalab submission file

python evalho3dv2.py --resume releasemodels/ho3dv2/realonly/checkpoint_200.pth --val_split test --json_folder jsonres/res

This will create a file 'pred.zip' ready for upload to the codalab challenge

• Download the HO3D-v2 dataset.

• launch training using python trainmeshreg and providing all arguments as in releasemodels/ho3dv2/realonly/opt.txt

Run the demo on the FPHAB dataset.

python visualize.py

This script loads three models and visualizes their predictions on samples from the test split of FPHAB:

• a model trained on the full FPHAB dataset
• a model trained with only a fraction (<1%) of the full ground truth annotations finetuned with photometric consistency
• a control model trained with the same fraction of the full ground truth annotations finetuned without photometric consistency

It produces images such as the following:

Run the training code

Train baseline model of entire FPHAB (100% of the data is supervised with 3D annotations)

python trainmeshreg.py --freeze_batchnorm --workers 8 --block_rot

• Step 1: Train baseline model on a fraction of the FPHAB dataset (here 0.65%)

python trainmeshreg.py --freeze_batchnorm --workers 8 --fraction 0.00625 --eval_freq 50

• Step 2: Resume training, adding photometric supervision

Step 1 will have produced a trained model which will be saved in a subdirectory of checkpoints/fhbhands_train_mini1/{data_you_launched_trainings}/.

Step 2 will resume training from this model, and further train with the additional photometric consistency loss on the frames for which the ground truth annotations are not used.

python trainmeshwarp.py --freeze_batchnorm --consist_gt_refs --workers 8 --fraction 0.00625 --resume checkpoints/path/to/saved/checkpoint.pth

• Optional: For fair comparison (same number of training epochs), training can also be resumed without photometric consistency (this shows that the improvement does not come simply from longer training)

python trainmeshwarp.py --freeze_batchnorm --consist_gt_refs --workers 8 --fraction 0.00625 --resume checkpoints/path/to/saved/checkpoint.pth --lambda_data 1 --lambda_consist 0

If you find this code useful for your research, consider citing our paper:

@INPROCEEDINGS{hasson20_handobjectconsist,
	       title     = {Leveraging Photometric Consistency over Time for Sparsely Supervised Hand-Object Reconstruction},
	       author    = {Hasson, Yana and Tekin, Bugra and Bogo, Federica and Laptev, Ivan and Pollefeys, Marc and Schmid, Cordelia},
	       booktitle = {CVPR},
	       year      = {2020}
}

Thanks to Samira Kaviani for spotting that in Table 2. the splits are different because I previously filtered out frames for which hands are further than 10cm away from the object ! I will rerun the results beginning September and update them here.

For this project, we relied on research code from:

• PWC-Net for image warping utilities.
• Christian Zimmermann for hand evaluation code from hand3d
• the PyTorch port of Neural Renderer.

I would like to specially thank Shreyas Hampali for advice on the HO-3D dataset and Guillermo Garcia-Hernando for advice and on the FPHAB dataset.

I would also like to thank Mihai Dusmanu, Yann Labbé and Thomas Eboli for helpful discussions and proofreading !