Erekle Shishniashvili

I developed a first neural network based Georgian text to speech algorithm. With my friends, we co-founded a project to gather five to six hours of pure Georgian audio without noise as a dataset and then use the state-of-the-art neural networks to train the model. This approach worked so well that it got the attention of all the big companies in the country as it was the first accurate real-time robust working Georgian text to speech system.

When the style transfer of painting styles to normal images through neural networks became famous, it also captured my attention. But in that period, style transfer only transferred the style of the image, not the artist itself. So, I decided to gather a lot of digital paintings of Georgian painters and modify the neural network to learn the style of the entire painter itself. This approach worked very well by looking at the style-transferred images gathered in one place and training for a long time. And I created a neural network that could learn the entire style of a painter. This approach was also better studied in academia. Still, I developed a web pipeline that generated an uploaded image in the style of different Georgian painters, allowing users to see how those painters would have drawn their photos. It was completely free, and more than a thousand people used it in the first 24 hours of its release.

During my thesis, I developed real-time efficient camera localization using both relative (calculated change of pose between images) and absolute (calculated 3D pose based on a single image using neural networks) regression methods. Apart from these two approaches I developed, a significant improvement was eliminating the need for labeled data. For my work method, we needed only video or images of the scene. To achieve this, I utilized the power of one of the variants of neural radiance fields to study the 3D geometry of the object without having labels for poses. Then I rendered synthetic images with artificial labels, generated smooth transitions between different angles of renderings, and trained the absolute pose or relative pose regression network.