LG AI Research: Introducing CVPR 2024 Best Paper and Low-level Vision Research

At CVPR 2024, a prestigious conference in the computer vision field, explosive interest in image and video synthesis was confirmed. In particular, both of the Best Papers selected at this conference were related to image/video synthesis research.

The Generative Image Dynamics research is a technology that predicts natural motion from a still image. Unlike existing research, it analyzes images in the Frequency domain, and is designed so that a Diffusion model predicts a Spectral volume. Through this, it generates natural, repeating videos such as trees or flowers swaying in the wind, and also provides an interactive demo that responds to a user's clicks or drags.

Meanwhile, LG AI Research presented Learning to Control Camera Exposure via Reinforcement Learning, a study in the Low-level Vision field. This research addresses Camera Exposure Control, the stage for acquiring optimal images before computer vision algorithms operate.

Low-quality images caused by dark environments, backlighting, fast motion, and so on degrade the performance of subsequent vision tasks. This research aims to lay the technical foundation for securing high-quality images across various environments by controlling camera exposure using Reinforcement Learning.