dumyy / CPF

CPF: Learning a Contact Potential Field to Model the Hand-object Interaction

Contact Potential Field

This repo contains model, demo, and test codes of our paper: CPF: Learning a Contact Potential Field to Model the Hand-object Interaction

Guide to the Demo

1. Get our code:

$ git clone --recursive https://github.com/lixiny/CPF.git
$ cd CPF

2. Set up your new environment:

$ conda env create -f environment.yaml
$ conda activate cpf

3. Download assets files and put it in assets folder.

Download the MANO model files from official MANO website, and put it into assets/mano. We currently only use the MANO_RIGHT.pkl

Now your assets folder should look like this:

.
├── anchor/
│   ├── anchor_mapping_path.pkl
│   ├── anchor_weight.txt
│   ├── face_vertex_idx.txt
│   └── merged_vertex_assignment.txt
├── closed_hand/
│   └── hand_mesh_close.obj
├── fhbhands_fits/
│   ├── Subject_1/
│   │   ├── ...
│   ├── Subject_2/
|   ├── ...
├── hand_palm_full.txt
└── mano/
    ├── fhb_skel_centeridx9.pkl
    ├── info.txt
    ├── LICENSE.txt
    └── MANO_RIGHT.pkl

4. Download Dataset

First-Person Hand Action Benchmark (fhb)

Download and unzip the First-Person Hand Action Benchmark dataset following the official instructions to the data/fhbhands folder If everything is correct, your data/fhbhands should look like this:

.
├── action_object_info.txt
├── action_sequences_normalized/
├── change_log.txt
├── data_split_action_recognition.txt
├── file_system.jpg
├── Hand_pose_annotation_v1/
├── Object_6D_pose_annotation_v1_1/
├── Object_models/
├── Subjects_info/
├── Video_files/
├── Video_files_480/ # Optionally

Optionally, resize the images (speeds up training !) based on the handobjectconsist/reduce_fphab.py.

$ python reduce_fphab.py

Download our fhbhands_supp and place it at data/fhbhands_supp:

Download our fhbhands_example and place it at data/fhbhands_example. This fhbhands_example contains 10 samples that are designed to demonstrate our pipeline.

├── fhbhands/
├── fhbhands_supp/
│   ├── Object_models/
│   └── Object_models_binvox/
├── fhbhands_example/
│   ├── annotations/
│   ├── images/
│   ├── object_models/
│   └── sample_list.txt

HO3D

Download and unzip the HO3D dataset following the official instructions to the data/HO3D folder. if everything is correct, the HO3D & YCB folder in your data should look like this:

data/
├── HO3D/
│   ├── evaluation/
│   ├── evaluation.txt
│   ├── train/
│   └── train.txt
├── YCB_models/
│   ├── 002_master_chef_can/
│   ├── ...

Download our YCB_models_supp and place it at data/YCB_models_supp

Now the data folder should have a root structure like:

data/
├── fhbhands/
├── fhbhands_supp/
├── fhbhands_example/
├── HO3D/
├── YCB_models/
├── YCB_models_supp/

5. Download pre-trained checkpoints

download our pre-trained CPF_checkpoints, unzip it at the CPF_checkpoints folder:

CPF_checkpoints/
├── honet/
│   ├── fhb/
│   ├── ho3dofficial/
│   └── ho3dv1/
├── picr/
│   ├── fhb/
│   ├── ho3dofficial/
│   └── ho3dv1/

6. Launch visualization

We create a FHBExample dataset in hocontact/hodatasets/fhb_example.py that only contains 10 samples to demonstrate our pipeline. Notice: this demo requires active screen for visualizing. Press q in the "runtime hand" window to start fitting.

$ python training/run_demo.py \
    --gpu 0 \
    --init_ckpt CPF_checkpoints/picr/fhb/checkpoint_200.pth.tar \
    --honet_mano_fhb_hand

7. Test on full dataset (FHB, HO3D v1/v2)

We provide shell srcipts to test on the full dataset to approximately reproduce our results.

FHB

dump the results of HoNet and PiCR:

$ python training/dumppicr_dist.py \
    --gpu 0,1 \
    --dist_master_addr localhost \
    --dist_master_port 12355 \
    --exp_keyword fhb \
    --train_datasets fhb \
    --train_splits train \
    --val_dataset fhb \
    --val_split test \
    --split_mode actions \
    --batch_size 8 \
    --dump_eval \
    --dump \
    --vertex_contact_thresh 0.8 \
    --filter_thresh 5.0 \
    --dump_prefix common/picr \
    --init_ckpt CPF_checkpoints/picr/fhb/checkpoint_200.pth.tar

and reload the GeO optimizer:

# setting 1: hand-only
$ CUDA_VISIBLE_DEVICES=0,1,2,3 python training/optimize.py \
    --n_workers 16 \
    --data_path common/picr/fhbhands/test_actions_mf1.0_rf0.25_fct5.0_ec \
    --mode hand

# setting 2: hand-obj
$ CUDA_VISIBLE_DEVICES=0,1,2,3 python training/optimize.py \
    --n_workers 16 \
    --data_path common/picr/fhbhands/test_actions_mf1.0_rf0.25_fct5.0_ec \
    --mode hand_obj \
    --compensate_tsl

HO3Dv1

dump:

$ python training/dumppicr_dist.py  \
    --gpu 0,1 \
    --dist_master_addr localhost \
    --dist_master_port 12356 \
    --exp_keyword ho3dv1 \
    --train_datasets ho3d \
    --train_splits train \
    --val_dataset ho3d \
    --val_split test \
    --split_mode objects \
    --batch_size 4 \
    --dump_eval \
    --dump \
    --vertex_contact_thresh 0.8 \
    --filter_thresh 5.0 \
    --dump_prefix common/picr_ho3dv1 \
    --init_ckpt CPF_checkpoints/picr/ho3dv1/checkpoint_300.pth.tar

and reload optimizer:

# hand-only
$ CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python training/optimize.py \
    --n_workers 24 \
    --data_path common/picr_ho3dv1/HO3D/test_objects_mf1_likev1_fct5.0_ec/ \
    --lr 1e-2 \
    --n_iter 500 \
    --hodata_no_use_cache \
    --lambda_contact_loss 10.0 \
    --lambda_repulsion_loss 4.0 \
    --repulsion_query 0.030 \
    --repulsion_threshold 0.080 \
    --mode hand

# hand-obj
$ CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python training/optimize.py \
    --n_workers 24 \
    --data_path common/picr_ho3dv1/HO3D/test_objects_mf1_likev1_fct5.0_ec/ \
    --lr 1e-2 \
    --n_iter 500  \
    --hodata_no_use_cache \
    --lambda_contact_loss 10.0 \
    --lambda_repulsion_loss 6.0 \
    --repulsion_query 0.030 \
    --repulsion_threshold 0.080 \
    --mode hand_obj

HO3Dofficial

dump:

$ python training/dumppicr_dist.py  \
    --gpu 0,1 \
    --dist_master_addr localhost \
    --dist_master_port 12356 \
    --exp_keyword ho3dofficial \
    --train_datasets ho3d \
    --train_splits val \
    --val_dataset ho3d \
    --val_split test \
    --split_mode official \
    --batch_size 4 \
    --dump_eval \
    --dump \
    --test_dump \
    --vertex_contact_thresh 0.8 \
    --filter_thresh 5.0 \
    --dump_prefix common/picr_ho3dofficial \
    --init_ckpt CPF_checkpoints/picr/ho3dofficial/checkpoint_300.pth.tar

and reload optimizer:

$ CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python training/optimize.py \
    --n_workers 24 \
    --data_path common/picr_ho3dofficial/HO3D/test_official_mf1_likev1_fct\(x\)_ec/  \
    --lr 1e-2 \
    --n_iter 500 \
    --hodata_no_use_cache \
    --lambda_contact_loss 10.0 \
    --lambda_repulsion_loss 2.0 \
    --repulsion_query 0.030 \
    --repulsion_threshold 0.080 \
    --mode hand_obj

Results

Testing on the full dataset may take a while ( 0.5 ~ 1.5 day ), thus we also provide our test results at fitting_res.txt.

K-MANO

We provide pytorch implementation of our Kinematic-chained MANO in lixiny/manopth, which is modified from the original hassony2/manopth. Thank Yana Hasson for providing the code.

Citation

If you find this work helpful, please consider citing us:

@article{yang2020cpf,
  title={CPF: Learning a Contact Potential Field to Model the Hand-object Interaction},
  author={Yang, Lixin and Zhan, Xinyu and Li, Kailin and Xu, Wenqiang and Li, Jiefeng and Lu, Cewu},
  journal={arXiv preprint arXiv:2012.00924},
  year={2020}
}

And if you have any question or suggestion, do not hesitate to contact me through siriusyang[at]sjtu[dot]edu[dot]cn.