Personal video highlights, made on the viewer's own device
A viewer picks what they care about and gets a fast-forward summary of a long video, built on a small device. The video and the viewer's preferences never go to a server.
The challenge
Personalized highlights normally mean uploading what someone watches and likes to a server, then processing every frame of every video there. It's expensive to run and it hands over private viewing data.
What we built
- Instead of decoding the whole video, the device looks at the small preview images streaming services already make for their seek bars.
- A compact vision model spots the moments that match the viewer's preferences.
- The player downloads only those moments and plays them back as a personal fast-forward summary.
Results
- Runs entirely on an NVIDIA Jetson Nano, a credit-card-sized computer; shown on movies, sports and TV shows.
- Demonstrated live at CVPR 2025, the leading computer vision conference.
- Related U.S. patent application published (US 2024/0073493 A1).
Where this applies
- Streaming and sports apps that want personal highlights without new servers
- Set-top boxes, smart TVs and other low-power devices
- Any product where viewing data should stay private
Under the hood
- Input is thumbnail containers (seek-bar preview sprites) rather than decoded frames.
- A lightweight 2D CNN classifies preferred content and returns timestamps; only the matching HLS segments are fetched.
- Paper - Mujtaba and Ryu, "EdgeVidSum - Real-Time Personalized Video Summarization at the Edge," CVPR 2025 Demo.
PyTorch2D CNNHLSNVIDIA Jetson Nano
