Jing Lin, Ailing Zeng, Haoqian Wang, Lei Zhang, Yu Li

The proposed UBody dataset

• 2024.08.26 : Update the implementation of the re-projection from SMPL-X to whole-body 2d keypoints (e.g., to align the 3D-to-2D keypoints), please check 3. Quick demo [Update information], Thanks to Yuhang Yang.
• 2023.10.12 : UBody is now supported in MMPose. Please feel free to use it. 🌟
• 2023.07.28 : UBody can boost 2D whole-body pose estimation and controllable image generation, especially for in-the-wild hand keypoint detection. The training and test code and pre-trained models are released. See details. 🥳
• 2023.05.03 : UBody-V1 is released. We'll release UBody-V2 later, which have manually annotated bboxes. 🕺
• 2023.04.17 : We fix bug of rendering in A100/V100 and support yolov5 as a person detector in demo.py. 🚀
• 2023.04.15 : We merge OSX into Grounded-SAM and support promptable 3D whole-body mesh recovery. 🔥

Demo of Grounded-SAM-OSX.

A person with pink clothes

A man with a sunglasses

This repo is official PyTorch implementation of One-Stage 3D Whole-Body Mesh Recovery with Component Aware Transformer (CVPR2023). We propose the first one-stage whole-body mesh recovery method (OSX) and build a large-scale upper-body dataset (UBody). It is the top-1 method on AGORA benchmark SMPL-X Leaderboard (dated March 2023).

• PyTorch >= 1.7 + CUDA

  Recommend to install by:

  pip install torch==1.11.0+cu113 torchvision==0.12.0+cu113 torchaudio==0.11.0 --extra-index-url https://download.pytorch.org/whl/cu113

• Python packages:

  bash install.sh

• Download the pre-trained OSX from here.
• Prepare pre-trained snapshot at pretrained_models folder.
• Prepare human_model_files folder following below Directory part and place it at common/utils/human_model_files.
• Go to demo folders, and run python demo.py --gpu 0 --img_path IMG_PATH --output_folder OUTPUT_FOLDER . Please replace IMG_PATH and OUTPUT_FOLDRE with your own image path and saving folder. For a more efficient inference, you can add --decoder_setting wo_decoder --pretrained_model_path ../pretrained_models/osx_l_wo_decoder.pth.tar to use the encoder-only version OSX.
• [Update information] The inference code will output the projected 2d kpts with shape (137,2), please refer to here for details of these 137 key points. The SMPLX version we use has 144 joints, please refer to this line. if you want to use COCO format key points, please refer to here. Note: the key points projected onto the image may be misaligned with humans, this is likely due to inaccurate boxes provided by detection models. It is recommended to use more advanced detection models or manually set bounding boxes.
• If you run this code in ssh environment without display device, do follow:

1、Install oemesa follow https://pyrender.readthedocs.io/en/latest/install/
2、Reinstall the specific pyopengl fork: https://github.com/mmatl/pyopengl
3、Set opengl's backend to egl or osmesa via os.environ["PYOPENGL_PLATFORM"] = "egl"

The ${ROOT} is described as below.

${ROOT}  
|-- data  
|-- dataset
|-- demo
|-- main  
|-- pretrained_models
|-- tool
|-- output  
|-- common
|   |-- utils
|   |   |-- human_model_files
|   |   |   |-- smpl
|   |   |   |   |-- SMPL_NEUTRAL.pkl
|   |   |   |   |-- SMPL_MALE.pkl
|   |   |   |   |-- SMPL_FEMALE.pkl
|   |   |   |-- smplx
|   |   |   |   |-- MANO_SMPLX_vertex_ids.pkl
|   |   |   |   |-- SMPL-X__FLAME_vertex_ids.npy
|   |   |   |   |-- SMPLX_NEUTRAL.pkl
|   |   |   |   |-- SMPLX_to_J14.pkl
|   |   |   |   |-- SMPLX_NEUTRAL.npz
|   |   |   |   |-- SMPLX_MALE.npz
|   |   |   |   |-- SMPLX_FEMALE.npz
|   |   |   |-- mano
|   |   |   |   |-- MANO_LEFT.pkl
|   |   |   |   |-- MANO_RIGHT.pkl
|   |   |   |-- flame
|   |   |   |   |-- flame_dynamic_embedding.npy
|   |   |   |   |-- flame_static_embedding.pkl
|   |   |   |   |-- FLAME_NEUTRAL.pkl

• data contains data loading codes.
• dataset contains soft links to images and annotations directories.
• pretrained_models contains pretrained models.
• demo contains demo codes.
• main contains high-level codes for training or testing the network.
• tool contains pre-processing codes of AGORA and pytorch model editing codes.
• output contains log, trained models, visualized outputs, and test result.
• common contains kernel codes for Hand4Whole.
• human_model_files contains smpl, smplx, mano, and flame 3D model files. Download the files from [smpl] [smplx] [SMPLX_to_J14.pkl] [mano] [flame]. We provide the download links for each file here.

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

${ROOT}  
|-- dataset  
|   |-- AGORA
|   |   |-- data
|   |   |   |-- AGORA_train.json
|   |   |   |-- AGORA_validation.json
|   |   |   |-- AGORA_test_bbox.json
|   |   |   |-- 1280x720
|   |   |   |-- 3840x2160
|   |-- EHF
|   |   |-- data
|   |   |   |-- EHF.json
|   |-- Human36M  
|   |   |-- images  
|   |   |-- annotations  
|   |-- MPII
|   |   |-- data
|   |   |   |-- images
|   |   |   |-- annotations
|   |-- MPI_INF_3DHP
|   |   |-- data
|   |   |   |-- images_1k
|   |   |   |-- MPI-INF-3DHP_1k.json
|   |   |   |-- MPI-INF-3DHP_camera_1k.json
|   |   |   |-- MPI-INF-3DHP_joint_3d.json
|   |   |   |-- MPI-INF-3DHP_SMPL_NeuralAnnot.json
|   |-- MSCOCO  
|   |   |-- images  
|   |   |   |-- train2017  
|   |   |   |-- val2017  
|   |   |-- annotations 
|   |-- PW3D
|   |   |-- data
|   |   |   |-- 3DPW_train.json
|   |   |   |-- 3DPW_validation.json
|   |   |   |-- 3DPW_test.json
|   |   |-- imageFiles
|   |-- UBody
|   |   |-- images
|   |   |-- videos
|   |   |-- annotations
|   |   |-- splits
|   |   |   |-- inter_scene_test_list.npy
|   |   |   |-- intra_scene_test_list.npy

• Download AGORA parsed data [data][parsing codes]

• Download EHF parsed data [data]

• Download Human3.6M parsed data and SMPL-X parameters [data][SMPL-X parameters from NeuralAnnot]

• Download MPII parsed data and SMPL-X parameters [data][SMPL-X parameters from NeuralAnnot]

• Download MPI-INF-3DHP parsed data and SMPL-X parameters [data][SMPL-X parameters from NeuralAnnot]

• Download MSCOCO data and SMPL-X parameters [data][SMPL-X parameters]

• Download 3DPW parsed data [data]

• Download UBody dataset from [data] and run the following commond to convert the videos into images:

  cd tool/UBody
  python video2image.py
  

• All annotation files follow MSCOCO format. If you want to add your own dataset, you have to convert it to MSCOCO format.

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

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

• Creating output folder as soft link form is recommended instead of folder form because it would take large storage capacity.
• 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.

Download pretrained encoder osx_vit_l.pth and osx_vit_b.pth from here and place the pretrained model to pretrained_models/.

In the main folder, run

python train.py --gpu 0,1,2,3 --lr 1e-4 --exp_name output/train_setting1 --end_epoch 14 --train_batch_size 16

After training, run the following command to evaluate your pretrained model on EHF and AGORA-val:

# test on EHF
python test.py --gpu 0,1,2,3 --exp_name output/train_setting1/ --pretrained_model_path ../output/train_setting1/model_dump/snapshot_13.pth.tar --testset EHF
# test on AGORA-val
python test.py --gpu 0,1,2,3 --exp_name output/train_setting1/ --pretrained_model_path ../output/train_setting1/model_dump/snapshot_13.pth.tar --testset AGORA

To speed up, you can use a light-weight version OSX by change the encoder setting by adding --encoder_setting osx_b or change the decoder setting by adding --decoder_setting wo_face_decoder. We recommend adding --decoder_setting wo_face_decoder as it would obviously speed up and would not lead to significant performance decline. It takes about 20 hours to finish the training with one NVIDIA A100.

In the main folder, run

python train.py --gpu 0,1,2,3 --lr 1e-4 --exp_name output/train_setting2 --end_epoch 140 --train_batch_size 16  --agora_benchmark --decoder_setting wo_decoder

After training, run the following command to evaluate your pretrained model on AGORA-test:

python test.py --gpu 0,1,2,3 --exp_name output/train_setting2/ --pretrained_model_path ../output/train_setting2/model_dump/snapshot_139.pth.tar --testset AGORA --agora_benchmark --test_batch_size 64 --decoder_setting wo_decoder

The reconstruction result will be saved at output/train_setting2/result/.

You can zip the predictions folder into predictions.zip and submit it to the AGORA benchmark to obtain the evaluation metrics.

You can use a light-weight version OSX by adding --encoder_setting osx_b.

In the main folder, run

python train.py --gpu 0,1,2,3 --lr 1e-4 --exp_name output/train_setting3 --train_batch_size 16  --ubody_benchmark --decoder_setting wo_decoder

After training, run the following command to evaluate your pretrained model on UBody-test:

python test.py --gpu 0,1,2,3 --exp_name output/train_setting3/ --pretrained_model_path ../output/train_setting3/model_dump/snapshot_13.pth --testset UBody --test_batch_size 64 --decoder_setting wo_decoder 

The reconstruction result will be saved at output/train_setting3/result/.

Download pretrained models osx_l.pth.tar and osx_l_agora.pth.tar from here and place the pretrained model to pretrained_models/.

In the main folder, run

python test.py --gpu 0,1,2,3 --exp_name output/test_setting1 --pretrained_model_path ../pretrained_models/osx_l.pth.tar --testset EHF

In the main folder, run

python test.py --gpu 0,1,2,3 --exp_name output/test_setting1 --pretrained_model_path ../pretrained_models/osx_l.pth.tar --testset AGORA

In the main folder, run

python test.py --gpu 0,1,2,3 --exp_name output/test_setting2  --pretrained_model_path ../pretrained_models/osx_l_agora.pth.tar --testset AGORA --agora_benchmark --test_batch_size 64

The reconstruction result will be saved at output/test_setting2/result/.

You can zip the predictions folder into predictions.zip and submit it to the AGORA benchmark to obtain the evaluation metrics.

In the main folder, run

python test.py --gpu 0,1,2,3 --exp_name output/test_setting3  --pretrained_model_path ../pretrained_models/osx_l_wo_decoder.pth.tar --testset UBody --test_batch_size 64

The reconstruction result will be saved at output/test_setting3/result/.

• RuntimeError: Subtraction, the '-' operator, with a bool tensor is not supported. If you are trying to invert a mask, use the '~' or 'logical_not()' operator instead.: Go to here

• TypeError: startswith first arg must be bytes or a tuple of bytes, not str.: Go to here.

This repo is mainly based on Hand4Whole. We thank the well-organized code and patient answers of Gyeongsik Moon in the issue!

@inproceedings{lin2023one,
  title={One-Stage 3D Whole-Body Mesh Recovery with Component Aware Transformer},
  author={Lin, Jing and Zeng, Ailing and Wang, Haoqian and Zhang, Lei and Li, Yu},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={21159--21168},
  year={2023}
}