|Topic:||Facial Image Manipulation by CNNs|
|Supervisor:||Ing. Jan Čech Ph.D.|
Convolutional Neural Networks (CNN) have proven excellent performance in many face related computer vision problems, often outperforming an average human annotator, e.g. Face recognition, biological age estimation.
The network is usually trained from a large dataset of images with annotated facial attributes (e.g. age, gender). When trained, the network takes an image as an input and predicts the facial attribute. Nevertheless, the network can be easily reverted: Given the attributes (and the initial image), find the image that would produce the given attributes. This problem is ambiguous, i.e. multiple images would provide the same attributes. Therefore, a regularization has to be employed to constrain the images to look like a natural image of a face. Recently, many successful facial generative models have been presented. This way, facial images can be manipulated - changing age, gender, identity, expressions of the subject on the image.
Your task will be to develop a demo-web application which would present the facial image manipulation. The input image will be uploaded, a user will set the attributes by scrollbars, the algorithm on the server will be executed and the output image will be presented.
We will provide the most of the manipulation algorithm: The trained network, a matlab prototype of the manipulation, and the generative model. Your task will be to test several design options and to integrate everything into a demo.
NVIDIA has recently released a pre-trained generative network that is able synthesize a photo-realistic high-resolution face images, see github. These results were widely noticed by a popular press, eg. here.
You will get in touch with the rapidly evolving deep convolutional network technology. The project workload can be naturally distributed among several team members. Impressive results are expected.
|Max.number of students:||0|