A Variational U-Net for Conditional Appearance and Shape Generation

This repository contains training code for the CVPR 2018 spotlight

A Variational U-Net for Conditional Appearance and Shape Generation

The model learns to infer appearance from a single image and can synthesize images with that appearance in different poses.

Project page with more results

Notes

This is a slightly modified version of the code that was used to produce the results in the paper. The original code was cleaned up, the data dependent weight initialization was made compatible with tensorflow >= 1.3.0 and a unified model between the datasets is used. You can find the original code and checkpoints online but if you want to use them, please keep in mind that:

• the original checkpoints are not compatible with the graphs defined in this repository. You must use the original code distributed with the checkpoints.
• the original code uses a data dependent weight initialization scheme which does not work with tensorflow >= 1.3.0. You should use tensorflow==1.2.1.
• the original code became a bit of a mess and we can no longer provide support for it.

Requirements

The code was developed with Python 3. Dependencies can be installed with

pip install -r requirements.txt

These requirements correspond to the dependency versions used to generate the pretrained models but other versions might work as well.

Training

Download and unpack the desired dataset. This results in a folder containing an index.p file. Either add a symbolic link named data pointing to the download directory or adjust the path to the index.p file in the <dataset>.yaml config file.

For convenience, you can also run

./download_data.sh <dataset> <store_dir>

which will perform the above steps automatically. <dataset> can be one of coco, deepfashion or market. To train the model, run

python main.py --config <dataset>.yaml

By default, images and checkpoints are saved to log/<current date>. To change the log directory and other options, see

python main.py -h

and the corresponding configuration file. To obtain images of optimal quality it is recommended to train for a second round with a loss based on Gram matrices. To do so run

python main.py --config <dataset>_retrain.yaml --retrain --checkpoint <path to checkpoint of first round>

Pretrained models

You can find pretrained models online.

Other Datasets

To be able to train the model on your own dataset you must provide a pickled dictionary with the following keys:

• joint_order: list indicating the order of joints.
• imgs: list of paths to images (relative to pickle file).
• train: list of booleans indicating if this image belongs to training split
• joints: list of [0,1] normalized xy joint coordinates of shape (len(joint_jorder), 2). Use negative values for occluded joints.

joint_order should contain

'rankle', 'rknee', 'rhip', 'rshoulder', 'relbow', 'rwrist', 'reye', 'lankle', 'lknee', 'lhip', 'lshoulder', 'lelbow', 'lwrist', 'leye', 'cnose'

and images without valid values for rhip, rshoulder, lhip, lshoulder are ignored.