Scientists from SAIC-M presented their projects at Samsung AI Forum 2021

Scientists from Samsung Artificial Intelligence Center – Moscow presented their projects at Samsung AI Forum 2021.
Samsung AI Forum, where academics and AI experts discuss the latest technological trends, future of artificial intelligence and research results, is held annually by Samsung Electronics. The event brings together leading experts in the field, world-renowned scientists and researchers as well as representatives of promising startups and big companies. In 2021 Samsung AI Forum was held as an online event and was broadcasted on Samsung Electronics YouTube channel.

The second day of the forum was organized by Samsung Research, Samsung Electronics’ advanced R&D hub. In his introduction speech, Dr. Sebastian Seung, President and Head of Samsung Research, presented an overview of the various AI-related projects, which Samsung Research is working on. He also talked about potential implementation of AI technologies in devices which will unveil new opportunities for self-expression and creativity, including new approach to using a smartphone camera as well as using a smartphone to manage other devices such as TV or air conditioner.

Among those promising AI technologies Dr. Sebastian Seung acknowledged LaMa (Resolution-robust Large Mask Inpainting with Fourier Convolutions) – a new approach to inpainting, a technology that removes objects from the picture while also making removal seamless for the user. Samsung AI Center – Moscow (SAIC-M) researchers created an algorithm which produces significantly more plausible result and is more robust in real-world applications than other existing technologies in the field. Gleb Sterkin, leader of Multimodal Data Analysis Lab in Moscow AI Center, has explained this and other projects in his presentation on AI Forum.

According to Gleb Sterkin, the objective of inpainting technology is to predict the background behind the object based on the context, i.e. wallpaper pattern or the texture of the wall etc., then remove the object and cover it in a realistic manner. In this projects his lab has a new approach to training such kind of models, which makes removal process more precise and robust across different image resolutions. The technology has all the components revisited – from training data generation to architecture and training objectives – compared to previous works.

Another project the team is working on is ‘scene relighting’ – a technology that allows demonstrating how land and cityscapes would look depending on time of day, lighting and weather conditions. The algorithm invented by researchers at Samsung AI Center – Moscow works with high resolution images up to 4K and produces extremely realistic results on landscape images.

Lastly, Gleb Sterkin presented ‘landscape animation’. It is a neural model which was trained on a few videos and many images to transform a single landscape image into a video: the water and the sky start moving smoothly on a resulting animation. A distinctive feature of this particular technology is that the user can set a preferable direction and intensity of such movements – and the method can generate infinite and never-repeating output. This could potentially be implemented in smartphones and TVs as an animated screen image.

There are seven Samsung Artificial Intelligence Centers around the world at the moment: Silicon Valley, New-York, Cambridge, Moscow, Toronto and Montreal as well as a Global AI Center in Seoul.

Full recording of the conference can be found on Samsung Elecrtonics’ YouTube channel.