Awesome

3D Human Mesh Recovery with Sequentially Global Rotation Estimation

[Paper]

3D Human Mesh Recovery with Sequentially Global Rotation Estimation
Dongkai Wang, Shiliang Zhang
ICCV 2023

Installation

1. Clone code

    git clone https://github.com/kennethwdk/SGRE
    cd ./SGRE

2. Create a conda environment for this repo

    conda create -n SGRE python=3.9
    conda activate SGRE

3. Install PyTorch >= 1.6.0 following official instruction, e.g.,

    conda install pytorch torchvision cudatoolkit=11.3 -c pytorch

We conduct experiments based on PyTorch 1.11.0 and cudatoolkit 11.3.1. You can follow this setting to reproduce our results.

4. Install other dependency python packages

    pip install -r requirements.txt

You should slightly change torchgeometry kernel code following here.

5. Prepare dataset

Please follow 3DCrowdNet to preparee dataset, and they should look like the following structure.

${ROOT}  
|-- data 
|   |-- J_regressor_extra.npy 
|   |-- Human36M  
|   |   |-- images  
|   |   |-- annotations   
|   |   |-- J_regressor_h36m_correct.npy
|   |-- MuCo  
|   |   |-- data  
|   |   |   |-- augmented_set  
|   |   |   |-- unaugmented_set  
|   |   |   |-- MuCo-3DHP.json
|   |   |   |-- smpl_param.json
|   |-- MSCOCO  
|   |   |-- images  
|   |   |   |-- train2017  
|   |   |   |-- val2017  
|   |   |-- annotations  
|   |   |-- J_regressor_coco_hip_smpl.npy
|   |-- MPII
|   |   |-- data  
|   |   |   |-- annotations
|   |   |   |-- images
|   |-- PW3D
|   |   |-- data
|   |   |   |-- 3DPW_latest_train.json
|   |   |   |-- 3DPW_latest_validation.json
|   |   |   |-- 3DPW_latest_test.json
|   |   |   |-- 3DPW_validation_crowd_hhrnet_result.json
|   |   |   |-- imageFiles

Usage

1. Download trained model

• SGRE Model
• Pre-trained Model
• SMPL Model, you should download basicModel_f_lbs_10_207_0_v1.0.0.pkl, basicModel_m_lbs_10_207_0_v1.0.0.pkl, and basicModel_neutral_lbs_10_207_0_v1.0.0.pkl to ${ROOT}/common/utils/smplpytorch/smplpytorch/native/models.

2. Evaluate Model

In main folder run

# evaluate on 3DPW test set with 1 gpus
python test.py --gpu 0-0 --exp_dir ../output/exp_09-01_09_32 --test_epoch 10

3. Train Model

We use the pre-trained ResNet-50 weights on COCO. Download the file of weights from above pre-trained model link and place it under ${ROOT}/tool/.

In main folder run

python train.py --amp --continue --gpu 0-0

The experimental results are obtained by training on one NVIDIA RTX 3090.

Acknowledgement

The code is developed upon 3DCrowdNet. Many thanks to their contributions.