Ligthor is a system for extracting highlights from live streaming videos using implicit crowdsourcing. To reproduce our experiments, please refer to our repo.

Low Cost

Compared with deep learning, Lightor uses a small set of training data (1 video) and much simpler model to reach a high precision. The training time for Lightor is seconds while hours or even days for deep learning.

Good performance using single labeled video

High Generalization

Lightor adopts a general model and workflow, which performs well on many different domains of videos, while most of deep learning based approaches only work well on the domain of training data.

Precision of LOL model when applying on Dota2

Easy to Deploy

The user interface of Lightor is simply a video player with red dots on progress bar, which is very easy to deploy on existing live streaming platforms like Twitch.

Easy to deploy on existing platforms


Recorded Video

Lightor Interface

Extracted Highlights