zhiqiangdon / pose-adv-aug

Code for "Jointly Optimize Data Augmentation and Network Training: Adversarial Data Augmentation in Human Pose Estimation" (CVPR 2018)
Apache License 2.0
84 stars 14 forks source link
adversarial-learning data-augmentation human-pose-estimation

Jointly Optimize Data Augmentation and Network Training: Adversarial Data Augmentation in Human Pose Estimation

Training code for the paper Jointly Optimize Data Augmentation and Network Training: Adversarial Data Augmentation in Human Pose Estimation, CVPR 2018

Overview

Traditional random augmentation has two limitations. It doesn't consider the individual difference of training samples when doing augmentation. And it is also independent of the training status of the target network. To tackle these problems, we design an agent to learn more effective data augmentation.

Adversarial Data Augmentation in Human Pose Estimation

We model the training process as an adversarial learning problem. The agent (generator), conditioning on the individual samples and network status, tries to generate ''hard'' augmentations for the target network. The target network (discriminator), on the other hand, tries to learn better from the augmentations.

Adversarial Data Augmentation in Human Pose Estimation

Prerequisites

This package has the following requirements:

Installing

Install pytorch:

pip install http://download.pytorch.org/whl/cu90/torch-0.3.0.post4-cp27-cp27mu-linux_x86_64.whl

Install torchvision, scipy, matplotlib, dominate and visdom:

pip install torchvision scipy matplotlib dominate visdom

Training

The training is divided into three stages. First, we pretrain the pose network for 10 epochs. Then we use the fixed pose network to pretrain the augmentation agent. Finally, we jointly optimize these two.

1. Pretrain the Pose Network

python stack-hg.py --gpu_id 0 --exp_id stack-2-hgs --vis_env stack-2-hgs --is_train true --bs 24

2. Pretrain the Augmentation Agent

Use the pose network to collect the scale and rotation distributions to train the agent:

python collect-scale-ditri.py --gpu_id 0 --exp_id stack-2-hgs --load_prefix_pose lr-0.00025-10.pth.tar --bs 10
python collect-rotation-ditri.py --gpu_id 0 --exp_id stack-2-hgs --load_prefix_pose lr-0.00025-10.pth.tar --bs 10

Pretrain the agent:

python pretrain-s-r-agent.py --gpu_id 0 --exp_id stack-2-hgs --load_prefix_pose lr-0.00025-10.pth.tar --bs 24

3. Jointly Train the Pose Network and Agent

python joint-train-pose-s-r-agent.py --gpu_id 0 --exp_id stack-2-hgs --load_prefix_pose lr-0.00025-10.pth.tar --load_prefix_sr lr-0.00025-1.pth.tar --vis_env stack-2-hgs-joint --is_train true --bs 24 

Citation

If you find this code useful in your research, please consider citing:

@inproceedings{peng2018jointly,
  title={Jointly optimize data augmentation and network training: Adversarial data augmentation in human pose estimation},
  author={Peng, Xi and Tang, Zhiqiang and Yang, Fei and Feris, Rogerio S and Metaxas, Dimitris},
  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
  pages={2226--2234},
  year={2018}
}