This project is working in progress.
(Image above is just a concept. Basic idea follows Few-shot Face Reenactment task's method. See reference papers below.)
Myeong Gyu Lee et al, "Anti-motion Blur Method using Conditional Adversarial Networks", CSA2018
Research by using conditional GAN to remove image motion blur (Obtain better quality images with low-end camera)
In general, taking a high-speed object such as projectiles with a low-speed camera are accompanied by artifacts called so-called Motion blur. Motion blur is a phenomenon that the boundaries of a moving object diffuse unclearly. Motion blur is divided into a ‘captured motion blur’ and ‘display motion blur’. The formal occurs when the object moves faster than the camera shutter speed, and the later occurs due to the limitations of the display. In this study, we focus on the captured motive blur caused by the shutter speed of the camera. Generally, leveraging expensive high-speed camera equipment or using a de-blurring algorithm has been proposed to remove this type of blur. However high-speed cameras are too costly for the End-user to use, and de-blur algorithms have a problem that it takes quite a while to get remarkable results. Therefore we propose a method that uses a machine learning technique to obtain clear images even in low-end single RGB cameras with low frame rate.
Naver AI Vision Hackathon (Image retrieval challenge) Finalist (Participated in Network optimization, Hyperparameter tuning, Reference paper collection)
See demo footage here : IRMarker Generation Method for Dynamic Projection Mapping
see Demo footage here : [RFTracker]