Learning to Perceive and Control Time in Videos — Speed Detection, Slow-Mo Curation, and Temporal Super-Resolution

Self-supervised models detect speed changes and estimate playback speed by exploiting multimodal cues and temporal structure in videos. The learned models curate the largest slow-motion video dataset from noisy in-the-wild sources, enabling speed-conditioned video generation and temporal super-resolution that transforms low-FPS blurry videos into high-FPS sequences. This work treats time as a learnable, manipulable visual dimension, with applications in temporal forensics and controllable video generation.

• #video understanding
• #self-supervised learning
• #temporal reasoning
• #slow-motion
• #temporal super-resolution