Training CamStyle with CycleGAN and VAE for reid task

CamStyle is trained with CycleGAN-pytorch

Replace VAE as a generator based on zhongzhun007's work CamStyle

Preparation

Requirements: Python=3.6 and Pytorch=0.4.0

1. Install Pytorch

2. Download dataset

   • Market-1501 [BaiduYun] [GoogleDriver]

   • DukeMTMC-reID [BaiduYun] (password: chu1) [GoogleDriver]

   • Move them to 'CamStyle/CycleGAN-for-CamStyle/datasets/market (or duke)'

Train CamStyle models

# For Market-1501
sh train_market.sh
# For Duke
sh train_duke.sh

Generate CamStyle images

# For Market-1501
sh test_market.sh
# For Duke
sh test_duke.sh

Experiment result

Through FID (Frechet Inception Distance) and SSIM (Structural SIMilarity), two well-recognized indicators for evaluating image quality of GAN network , to campare the quality of images generated by Cycle-VAE-GAN and Cyle-GAN.

• Cycle-VAE-GAN training takes less time. Replacing the ResnetGenerator in the original paper with the VAE encoder greatly reduced the number of convolution layers and improved training time.
• Regardless of the visual contrast or FID, SSIM and other image quality indicators can be reflected, Cycle-VAE-GAN has better image generation capabilities.
• Compared with the original paper, the accuracy is better on its baseline, both mAP and Rank-1 are improved.

Conf

• Camera Style Adaptation for Person Re-identification
• Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networkss
• Image-to-Image Translation with Conditional Adversarial Networks