Pose-Guided Person Image Synthesis

This is mofied version of 'Pose-Guided Person Image Synthesis in the Non-iconic Views'. To run this code, you need to do as follows,

Prepare the environment

This code needs pytorch 1.4.0.

Download Datasets

We provide the processed frames, poses and segmentations for Penn Action and BBC-Pose. Market-1501 is the same as the one used in PATN.

Train

python main.py -n debug -R -nl -dp -d {Penn_Action/bbcpose/market1501} -dr {your_data_root}

Generate

python inference.py --dr {your_data_root} -n {name of model} -dp -nl

Then you will have the generated images in ./result_images/{dataset}/{name of model}.

Evaluate

We provide several metrics in ./metrics including Inception Score, FID, SSIM and M-SSIM.

You need to modify the variable dataset and root to your own setting, and run

python inception_score.py
python compute_ssim.py
python fid_score.py -c {gpu}

The scripts will evaluate the generated images from all models in the chosen dataset.