tfaehse
commented
6 months ago I also give my consent. But don't you think there would be a way to implement algorithms like BoT-SORT without changing the license? This would make it easier to use this repository in a commercial environment.
There would, I'm sure. But the integration in YOLOv8 is really amazingly well done, the additional tracking steps costs only a few percent in inference time for me, while accounting for motion compensation and visual similarity using the features from the detector: https://docs.ultralytics.com/modes/track/
In order to "fill in" for missed detections, especially running in backwards, I still want to run additional tracking steps that operate in a much simpler manner, but fully replicating the performance of the botsort integration looks fairly challenging. Additionally, the way I understand it, this project's ultralytics dependency can only be updated if we follow their license, and this will make it very difficult to stay up to date with python, PyTorch etc.
A while ago, ultralytics switched to AGPL 3.0: https://github.com/ultralytics/ultralytics/commit/2c6fc0a4443b9cf805ef17b1cfdd71a98693b4d4
As a result, we have to switch to this license as well if we want to use newer YOLOv8 features, like newer versions of the ultralytics package or using a new version to retrain some models. In particular, I would like to add the botsort algorithm natively supported in YOLOv8 now.
I need consent from all contributors here in order to "just" switch, so I will mention all of you here - you're not forced to agree of course, but I would like to ask you to review the implications (especially the need of any derivative work to be published under the same license) and give me some feedback. Thank you!
@joshinils @knrdl @dark-vex @Syperium @alexx-km