ERD-3DPose

by Yangyuxuan Kang, Anbang Yao, Shandong Wang, Ming Lu, Yurong Chen, Enhua Wu.

This project is the official implementation of "Explicit Residual Descent for 3D Human Pose Estimation from 2D Joint Locations", accepted by BMVC 2020. We provide Pytorch code and pretrained model for reproduction. Experiments are conducted on Human3.6M dataset.

Dependencies

• Python 2.7
• PyTorch >= 1.1.0

Installation

1. First, clone this repository:

   git clone https://github.com/kyang-06/ERD_3DPose.git
   cd ERD_3dpose
   

2. Install requirements:

   pip install -r requirements.txt
   

3. Download the pre-processed Human3.6M pose data (no image) and extract it into ./data:

   cd ./data/
   unzip ERD_data.zip
   

4. (Optional) Download pretrained model and extract it into ./checkpoint:

   cd ./checkpoint/
   unzip checkpoint.zip
   

Results

Here are the results of released pretrained model correspoding to Table 3 in the paper.

Lifting Network	Root Joint Locations	Weights&Error on GT2D	Weights&Error on DET2D	Note
SimpleBaseline [Martinez et al. ICCV'17]	GTRoot	ERD (baseline) : 32.6 (37.6)	ERD (baseline) : 50.0 (53.5)	Use PCA
LCN [Ci et al. ICCV'19]	GTRoot	ERD (baseline) : 35.4 (39.1)	ERD (baseline) : 50.2 (53.4)
SemGCN [Zhao et al. CVPR'19]	GTRoot	ERD (baseline) : 37.4 (40.7)	ERD (baseline) : 52.9 (56.6)

Experiment

Training lifting model

1. For different input source of 2D pose, change --input

python main.py --exp lifting_linear_gt --stage lifting  --lifting_model linear --input gt

python main.py --exp lifting_linear_cpn --stage lifting --lifting_model linear --input cpn

2. For different lifting model network structure, change --lifting_model

python main.py --exp lifting_lcn_gt --stage lifting --lifting_model lcn --knn 3 --hidsize 64 --input gt

python main.py --exp lifting_semgcn_gt --stage lifting --lifting_model semgcn --num_block 4 --hidsize 128 --input gt

Testing lifting model

python main.py --exp eval_lifting_linear_gt --test --lifting_model linear \
               --input gt --load checkpoint/lifting_linear_gt/ckpt_best.pth.tar

---

Training residual regressors

1. For different feature reconstruction type (Table 2 in the paper), change --inc_input_type

python main.py --exp ERD_linear_gt_proj --stage increment \
               --lifting_model linear --input gt \
               --load checkpoint/lifting_linear_gt/ckpt_best.pth.tar \
               --inc_input_type proj

python main.py --exp ERD_linear_gt_delta --stage increment \
               --lifting_model linear --input gt \
               --load checkpoint/lifting_linear_gt/ckpt_best.pth.tar \
               --inc_input_type delta

---

Testing residual regressors

python main.py --exp eval_ERD_linear_gt_delta --test --stage increment \
               --lifting_model linear --input gt \
               --load checkpoint/lifting_linear_gt/ckpt_best.pth.tar \
               --load_inc checkpoint/ERD_lifting_linear_gt/ckpt_best.pth.tar

---

Using PCA to normalize input data

When stage=lifting

python main.py --exp lifting_linear_gt_pca --stage lifting  --lifting_model linear \
               --input gt --pca_input --pca_component 30

When stage=increment

python main.py --exp ERD_linear_gt_pca_proj --stage increment \
               --lifting_model linear --input gt \
               --load checkpoint/lifting_linear_gt_pca/ckpt_best.pth.tar \
               --inc_input_type proj --pca_input --pca_component 30

Citing

If you find this code useful for your research or the use data generated by our method, please consider citing the following paper:

@Inproceedings{kang2020ERD,
  Title          = {Explicit Residual Descent for 3D Human Pose Estimation from 2D Joint Locations},
  Author         = {Yangyuxuan, Kang and Anbang, Yao and Shandong, Wang and Ming, Lu and Yurong, Chen and Enhua, Wu},
  Booktitle      = {BMVC},
  Year           = {2020}
}