using OpenCV and CNN (convolutional neural network)
To see project blog: https://adrianyi.com/2017/07/FacialFeatures.html
This is still going through implementation. Currently, you can run the following:
python FacialFilter.py
This will use the pre-trained model to add sunglasses to all faces in an image (example shown below).
Currently, this part uses Keras. I'll change it later to be purely in TensorFlow.
python train.py
This will call the TensorFlow meta graph and load pre-trained weights. It'll Iterate through 10 epochs with training data from Kaggle (you have to download it).
The code can handle re-initializing weights instead, but I have not implemented an argument parser.
See example below.
Requirements:
Download datasets from https://www.kaggle.com/c/facial-keypoints-detection/data and store the unzipped CSV files in data folder.
- Python 3.x
- OpenCV 3
- tensorflow 1.2.1
- pandas
Input:
Output:
This outputs something like this
Loading data...
Data loaded. Training...
Epoch: 1 - Loss: 0.00461 , Accuracy: 0.952
Valid Loss: 0.0044 , Valid Accuracy: 0.952
Epoch: 2 - Loss: 0.00444 , Accuracy: 0.951
Valid Loss: 0.00433 , Valid Accuracy: 0.952
Epoch: 3 - Loss: 0.00462 , Accuracy: 0.95
Valid Loss: 0.00437 , Valid Accuracy: 0.952
Epoch: 4 - Loss: 0.00443 , Accuracy: 0.952
Valid Loss: 0.00432 , Valid Accuracy: 0.952
Epoch: 5 - Loss: 0.00426 , Accuracy: 0.953
Valid Loss: 0.00433 , Valid Accuracy: 0.952
Epoch: 6 - Loss: 0.00459 , Accuracy: 0.951
Valid Loss: 0.00432 , Valid Accuracy: 0.952
Epoch: 7 - Loss: 0.00429 , Accuracy: 0.953
Valid Loss: 0.00433 , Valid Accuracy: 0.952
Epoch: 8 - Loss: 0.00431 , Accuracy: 0.952
Valid Loss: 0.00432 , Valid Accuracy: 0.952
Epoch: 9 - Loss: 0.00468 , Accuracy: 0.952
Valid Loss: 0.00432 , Valid Accuracy: 0.952
Epoch: 10 - Loss: 0.00416 , Accuracy: 0.953
Valid Loss: 0.00432 , Valid Accuracy: 0.952
Finished training. Testing it on an image.
Test image saved as test_facial_feature.jpg
where the test_facial_feature.jpg looks like this:
Not bad! :)