Simulated+Unsupervised (S+U) Learning in TensorFlow

TensorFlow implementation of Learning from Simulated and Unsupervised Images through Adversarial Training.

Requirements

• Python 2.7
• TensorFlow 0.12.1
• SciPy
• pillow
• tqdm

Usage

To generate synthetic dataset:

1. Run UnityEyes with changing resolution to 640x480 and Camera parameters to [0, 0, 20, 40].
2. Move generated images and json files into data/gaze/UnityEyes.

The data directory should looks like:

data
├── gaze
│   ├── MPIIGaze
│   │   └── Data
│   │       └── Normalized
│   │           ├── p00
│   │           ├── p01
│   │           └── ...
│   └── UnityEyes # contains images of UnityEyes
│       ├── 1.jpg
│       ├── 1.json
│       ├── 2.jpg
│       ├── 2.json
│       └── ...
├── __init__.py
├── gaze_data.py
├── hand_data.py
└── utils.py

To train a model (samples will be generated in samples directory):

$ python main.py
$ tensorboard --logdir=logs --host=0.0.0.0

To refine all synthetic images with a pretrained model:

$ python main.py --is_train=False --synthetic_image_dir="./data/gaze/UnityEyes/"

Training results

Differences with the paper

• Used Adam and Stochatstic Gradient Descent optimizer.
• Only used 83K (14% of 1.2M used by the paper) synthetic images from UnityEyes.
• Manually choose hyperparameters for B and lambda because those are not specified in the paper.

Experiments #1

For these synthetic images,

Result of lambda=1.0 with optimizer=sgd after 8,000 steps.

$ python main.py --reg_scale=1.0 --optimizer=sgd

Result of lambda=0.5 with optimizer=sgd after 8,000 steps.

$ python main.py --reg_scale=0.5 --optimizer=sgd

Training loss of discriminator and refiner when lambda is 1.0 (green) and 0.5 (yellow).

Experiments #2

For these synthetic images,

Result of lambda=1.0 with optimizer=adam after 4,000 steps.

$ python main.py --reg_scale=1.0 --optimizer=adam

Result of lambda=0.5 with optimizer=adam after 4,000 steps.

$ python main.py --reg_scale=0.5 --optimizer=adam

Result of lambda=0.1 with optimizer=adam after 4,000 steps.

$ python main.py --reg_scale=0.1 --optimizer=adam

Training loss of discriminator and refiner when lambda is 1.0 (blue), 0.5 (purple) and 0.1 (green).

Author

Taehoon Kim / @carpedm20