This talk presents research toward AI systems with similar capabilities through three connected stages

Abstract:
Humans can perceive, understand, and imagine dynamic environments. This talk presents research toward AI systems with similar capabilities through three connected stages: visual perception, understanding, and generation. At the perception stage, video restoration methods, including deraining and desnowing, recover reliable visual signals from degraded inputs. At the understanding stage, video detection and segmentation methods identify objects, maintain temporal consistency, and model entities and their relationships over time. At the generation stage, controllable generative models enable semantic video editing and synthesis from user instructions, allowing AI to imagine and create new visual scenarios. Together, these efforts aim to move machine intelligence from seeing visual signals clearly, to understanding dynamic scenes accurately, and ultimately to imagining and simulating possible outcomes in complex real-world and engineering environments.