• This repo is the official Pytorch implementation of Keypoint Transformer: Solving Joint Identification in Challenging Hands and Object Interactions for Accurate 3D Pose Estimation accepted at CVPR'22 (Oral)
• The dataset contains Evaluation and Training scripts for the network introduced in the paper using different datasets.

1. Setup the conda environment

   conda create --name kypt_trans python==3.8.11
   conda activate kypt_trans
   conda install pytorch==1.10.1 torchvision==0.11.2 torchaudio==0.10.1 cudatoolkit=11.3 -c pytorch -c conda-forge
   pip install -r requirements.txt
   

2. Download MANO model files from the website (requires login) and set the smplx_path in config.py

3. Clone the Current Repo

   git clone <curr_repo>
   cd kypt_trans
   cd main
   

The setup has been tested on NVIDIA 3090 GPU.

Depending on the dataset you intend to train/evaluate follow the instructions below for the setup.

1. Download the dataset from the website
2. In config.py, set interhand_anno_dir to point to the annotations directory
3. In config.py, set interhand_images_path to point to the images directory
4. If you intend to use RootNet output for the root joint translation, download the RootNet results for InterHand2.6M from here. Set root_net_output_path in config.py for point to the RootNet outputs folder. Instead, if you intend to test with ground-truth relative translation, set root_net_output_path to None

1. Download the dataset from the website and set ho3d_anno_dir in config.py to point to the dataset folder
2. Download the YCB models from here. The original mesh models are large and wont fit in the memory for some computation. Because of this the mesh decimated to 2000 faces as described here. Rename the decimated models as textured_simple_2000.obj

1. Download the dataset from the website and set h2o3d_anno_dir in config.py to point to the dataset folder.
2. Follow step 2 in HO-3D Setup above to download, decimate and set the object models.

We provide the demo script for visualizing the outputs.

1. Download the checkpoint file for the model trained to output MANO joint angles from here
2. Set dataset = 'InterHand2.6M' and pose_representation = 'angles' in config.py
3. Run the following script:

   python demo.py --ckpt_path <path_to_ckpt> --use_big_decoder --dec_layers 6
   

4. The outputs are shown in matplotlib and open3d windows. See the instructions in the command line to navigate

1. Download the checkpoint file for the model trained to output 3D pose representation from here
2. Set dataset = 'ho3d' and pose_representation = '3D' in config.py
3. Run the follwing script:

   python demo.py --ckpt_path <path_to_ckpt> --use_big_decoder --dec_layers 6
   

4. Since the output here is only 3D joint locations, the projections in 2D are shown in the matplotlib window. See the instructions in the command line to navigate.

Depending on the dataset you intend to evaluate follow the instructions below.

1. Make the following changes in the config.py

   dataset = 'InterHand2.6M'`
   pose_representation = '2p5D' # Table 1 in paper uses 2.5D pose representation
   

2. Download the checkpoint file from here

3. Run the following command:

   python test.py --ckpt_path <path_to_interhand2.6m_ckpt> --gpu_ids <gpu_ids>
   

   If running on multiple GPUs, set <gpu_ids> to 0,1,2,3

4. The error metrics are dumped into a .txt file in the folder containing the checkpoint

5. Final numbers as below:

[Results txt file]

1. Make the following changes in the config.py

   dataset = 'ho3d'
   pose_representation = '3D' # Table 2 in paper uses 3D pose representation
   

2. Download the checkpoint file from here
3. Run the following command:

   python test.py --ckpt_path <path_to_ho3d_ckpt> --use_big_decoder --dec_layers 6
   

4. The object error metric (MSSD) is dumped into a .txt file in the folder containing the checkpoint
5. Also dumped is a .json file which can be submitted to the HO-3D (v2) challenge after zipping the file
6. Here is the dumped results file after the run: [Results txt file]
7. Hand pose estimation accuracy in the HO-3D challenge leaderboard: here, user: bullet

1. Make the following changes in the config.py

   dataset = 'h2o3d'
   pose_representation = '3D' # h2o3d results in paper uses 3D pose representation
   

2. Download the checkpoint file from here
3. Run the following command:

   python test.py --ckpt_path <path_to_h2o3d_ckpt>
   

4. The object error metric (MSSD) is dumped into a .txt file in the folder containing the checkpoint
5. Also dumped is a .json file which can be submitted to the H₂O-3D challenge after zipping the file.
6. Here is the dumped results file after the run: [Results txt file]
7. Hand pose estimation accuracy in the H₂O-3D challenge leaderboard: here, user: bullet

1. Depending on the dataset and output pose respresentation you intend to train on, set the dataset & pose_representation variables in the config.py.
2. Run the following script to start the training:

   CUDA_VISIBLE_DEVICES=0,1 python train.py --run_dir_name <run_name>
   

   To continue training from the last saved checkpoint use --continue argument in the above command.
3. The checkpoints are dumped after every ecoch in the 'output' folder of the base directory
4. Tensorboard logging is also available in the 'output' folder

The H₂O-3D results in the paper are obtained by training the network on the combined dataset of HO-3D and H₂O-3D. This training can be achieved by setting dataset to ho3d_h2o3d in config.py.

    @InProceedings{Hampali_2022_CVPR_Kypt_Trans,  
    author = {Shreyas Hampali and Sayan Deb Sarkar and Mahdi Rad and Vincent Lepetit},  
    title = {Keypoint Transformer: Solving Joint Identification in Challenging Hands and Object Interactions for Accurate 3D Pose Estimation},  
    booktitle = {IEEE Computer Vision and Pattern Recognition Conference},  
    year = {2022}  
    }  

• A lot of the code has been reused from InterHand2.6M repo and the DETR repo. We thank the authors for making their code public