This repository is containing 200.000 raw images and 100.000 preprocessed 10 years challenge images and implementation details.
1 - You can use these datasets with providing the dataset publisher who is me :).
2 - Preprocessed data ( This dataset has problem that some old and young images are in wrong class. The users should change it. It needs crowd sourcing. )
1 - Opencv face detector used to detect faces in the raw images. --> Face Detector
2 - The images have two faces detected and the faces are cropped.
3 - Face alignment is applied that uses affine transformation and Dlib face landmark detection. --> Face Alignment
1 - 994 preprocessed images is selected.
2 - pix2pix model is used for creating 10 years changed image.
3 - After 200 epochs and approximately 10 hours( I know the training dataset small and training is limited due to that I used Google Colab as a resource for training. )
-
Due to high variability of faces that age and ethnicity difference, the head pose and glasses and other materials reconstructing the faces is challenging. I believe using head pose estimation will improve training results.
-
Since the dataset is collected from internet, therefore, the images that is almost 10 years old. The image quality is relatively low quality to process.
-
I will use more training images when I found time and computing power. Therefore, this repo will be updated...