• This repo is official PyTorch implementation of InterHand2.6M: A Dataset and Baseline for 3D Interacting Hand Pose Estimation from a Single RGB Image (ECCV 2020).
• Our InterHand2.6M dataset is the first large-scale real-captured dataset with accurate GT 3D interacting hand poses.

Above demo videos have low-quality frames because of the compression for the README upload.

• 2021.03.22. Finally, InterHand2.6M v1.0, which includes all images of 5 fps and 30 fps version, is released! 🎉
• 2020.11.26. Demo code for a random image is added! Checkout below instructions.
• 2020.11.26. Fitted MANO parameters are updated to the better ones (fitting error is about 5 mm). Also, reduced to much smaller file size by providing parameters fitted to the world coordinates (independent on the camera view).
• 2020.10.7. Fitted MANO parameters are available! They are obtained by NeuralAnnot.

• For the InterHand2.6M dataset download and instructions, go to [HOMEPAGE].
• Belows are instructions for our baseline model, InterNet, for 3D interacting hand pose estimation from a single RGB image.

1. Download pre-trained InterNet from here
2. Put the model at demo folder
3. Go to demo folder and edit bbox in here
4. run python demo.py --gpu 0 --test_epoch 20
5. You can see result_2D.jpg and 3D viewer.

1. Install SMPLX
2. cd tool/MANO_render
3. Set smplx_path in render.py
4. Run python render.py

1. cd tool/camera_visualize
2. Run python camera_visualize.py

• As there are many cameras, you'd better set subset and split in line 9 and 10, respectively, by yourself.

The ${ROOT} is described as below.

${ROOT}
|-- data
|-- common
|-- main
|-- output

• data contains data loading codes and soft links to images and annotations directories.
• common contains kernel codes for 3D interacting hand pose estimation.
• main contains high-level codes for training or testing the network.
• output contains log, trained models, visualized outputs, and test result.

You need to follow directory structure of the data as below.

${ROOT}
|-- data
|   |-- STB
|   |   |-- data
|   |   |-- rootnet_output
|   |   |   |-- rootnet_stb_output.json
|   |-- RHD
|   |   |-- data
|   |   |-- rootnet_output
|   |   |   |-- rootnet_rhd_output.json
|   |-- InterHand2.6M
|   |   |-- annotations
|   |   |   |-- all
|   |   |   |-- human_annot
|   |   |   |-- machine_annot
|   |   |-- images
|   |   |   |-- train
|   |   |   |-- val
|   |   |   |-- test
|   |   |-- rootnet_output
|   |   |   |-- rootnet_interhand2.6m_output_all_test.json
|   |   |   |-- rootnet_interhand2.6m_output_machine_annot_val.json

• Download InterHand2.6M data [HOMEPAGE]
• Download STB parsed data [images] [annotations]
• Download RHD parsed data [images] [annotations]
• All annotation files follow MS COCO format.
• If you want to add your own dataset, you have to convert it to MS COCO format.

You need to follow the directory structure of the output folder as below.

${ROOT}
|-- output
|   |-- log
|   |-- model_dump
|   |-- result
|   |-- vis

• log folder contains training log file.
• model_dump folder contains saved checkpoints for each epoch.
• result folder contains final estimation files generated in the testing stage.
• vis folder contains visualized results.

• In the main/config.py, you can change settings of the model including dataset to use and which root joint translation vector to use (from gt or from RootNet).

In the main folder, run

python train.py --gpu 0-3 --annot_subset $SUBSET

to train the network on the GPU 0,1,2,3. --gpu 0,1,2,3 can be used instead of --gpu 0-3. If you want to continue experiment, run use --continue.

$SUBSET is one of [all, human_annot, machine_annot].

• all: Combination of the human and machine annotation. Train (H+M) in the paper.
• human_annot: The human annotation. Train (H) in the paper.
• machine_annot: The machine annotation. Train (M) in the paper.

Place trained model at the output/model_dump/.

In the main folder, run

python test.py --gpu 0-3 --test_epoch 20 --test_set $DB_SPLIT --annot_subset $SUBSET

to test the network on the GPU 0,1,2,3 with snapshot_20.pth.tar. --gpu 0,1,2,3 can be used instead of --gpu 0-3.

$DB_SPLIT is one of [val,test].

• val: The validation set. Val in the paper.
• test: The test set. Test in the paper.

$SUBSET is one of [all, human_annot, machine_annot].

• all: Combination of the human and machine annotation. (H+M) in the paper.
• human_annot: The human annotation. (H) in the paper.
• machine_annot: The machine annotation. (M) in the paper.

Here I provide the performance and pre-trained snapshots of InterNet, and output of the RootNet as well.

• [Trained on InterHand2.6M 5 fps H+M]
• [Trained on InterHand2.6M 5 fps H]
• [Trained on InterHand2.6M 5 fps M]
• [Trained on STB]
• [Trained on RHD]

• [Output on InterHand2.6M]
• [Output on STB]
• [Output on RHD]

• Codes
• See RootNet for the code instructions.

@InProceedings{Moon_2020_ECCV_InterHand2.6M,  
author = {Moon, Gyeongsik and Yu, Shoou-I and Wen, He and Shiratori, Takaaki and Lee, Kyoung Mu},  
title = {InterHand2.6M: A Dataset and Baseline for 3D Interacting Hand Pose Estimation from a Single RGB Image},  
booktitle = {European Conference on Computer Vision (ECCV)},  
year = {2020}  
}  

InterHand2.6M is CC-BY-NC 4.0 licensed, as found in the LICENSE file.

[Terms of Use] [Privacy Policy]