ml-lab / pose-gan

Deformable GANs for Pose-based Human Image Generation.

Requirment

• python2
• Numpy
• Scipy
• Skimage
• Pandas
• Tensorflow
• Keras
• tqdm

Training

In orger to train a model:

1. Create folder market-dataset with 2 subfolder (test and train). Put the test images in test images in test/ and train images in train/.
2. Download pose estimator (conversion of this https://github.com/ZheC/Realtime_Multi-Person_Pose_Estimation) pose_estimator.h5. Launch python compute_cordinates.py. It will compute human keypoints.
3. Create pairs dataset with python create_pairs_dataset.py. It define pairs for training.
4. Run python train.py (see list of parameters in cmd.py)

Testing

0. Download checkpoints (https://yadi.sk/d/dxVvYxBw3QuUT9).
1. Run python test.py --generator_checkpoint path/to/generator/checkpoint (and same parameters as in train.py). It generate images and compute inception score, SSIM score and their masked versions.
2. To compute ssd_score. Download pretrained on VOC 300x300 model from https://github.com/weiliu89/caffe/tree/ssd. Put it in the ssd_score forlder. Run python compute_ssd_score.py --input_dir path/to/generated/images --img_index 2