We plan to explore manifold learning as presented in [@isomap] for a lower dimensional embedding of images. Since lower dimensional hidden representations allow for generative modeling, we also plan to explore image generation by sampling from this lower dimensional manifold.
We plan to go ahead with Celebface Attribute Dataset. For computational purposes, we randomly choose roughly 6000 images from a total set of more 200k images. To reduce the complexity, we convert all the images to grayscale.
We will begin with choosing an appropriate dataset from a plethora that is already available. We will then detect keypoints followed by generating descriptors around those keypoints using a suitable algorithm. Once we obtain the feature vector representation for each image, we will try to learn a parametrized manifold space. This parametrization would not only help us in generating new faces along that manifold, but a suitable choice of parameter may help us in morphing specific features of individual faces.
We were able to successfully generate the following videos :