Have you ever wondered how to make videos smaller without losing important details? Well, scientists have come up with a clever idea: using afterimages. Imagine, instead of saving each frame, you only keep the important bits that your eyes and brain combine to make a complete picture.

First, let's look at how videos change. Some scenes are simple, like a still landscape, while others are complex, like a crowded street. The method these clever folks used starts by figuring out which scenes are simple or complex. This helps decide how many keyframes (important frames) to keep.

Then, they use something called optical flow to track how objects move. This helps create masks that show where objects are moving. These masks are combined over time in a process called alpha blending. The result? A single afterimage that holds a lot of information from many frames.

Sounds interesting, right? They tested this on a dataset called UCF-Crime and found the videos were squeezed down to just 4. 03% of their original size. Even better, the compressed videos were almost as good as the originals in telling normal from abnormal behaviors. For multi-class classification, they even outperformed the originals. In fact, when focusing on just abnormal behaviors, the performance jumped by a significant 4. 25%.

But wait, there's more! Big language models (LLMs) could understand the story of the original videos just from these afterimages. That's pretty amazing, huh?

Video Crunch: Shrinking Videos by Seeing Past Frames

Actions