DETRs Beat YOLOs on Real-Time Object Detection

[jasondavies]

https://zhao-yian.github.io/RTDETR/

[isoprophlex]

Good to see some progress in the space, as NMS always felt like a very hacky solution to me.

Also good that the YOLO moniker is being challenged. After pjreddie went off to do better things, I've always felt a bit sad about random parties co-opting the YOLO name. And then with Ultralyics and their weird approach to monetizing YOLOv8 "yeah uh so it's open source but if you actually train a model plz fork over your money)... not a nice look.

[wongarsu → isoprophlex]

The upside of Ultralytic's monetization is that the tooling is top-notch. You really can fine tune a model in about three lines of code, train a model from scratch just as easily, and run inference with ease. And all with decent documentation.

For my use cases I prefer great AGPL (with optional commercial license) tooling over the free good-enough-for-the-paper-but-not-practical-applications tooling you get with most computer vision projects (including this one).

[eddiewithzato → isoprophlex]

thought he quit because didn’t want to become complacent in his work being used in wars, oppressive regimes, etc.

[hprotagonist → eddiewithzato]

yes. lack of that complicity seemed like “better things” to him, i think.

[isoprophlex → hprotagonist]

Exactly, the man's a hero in my book.

[hprotagonist → isoprophlex]

someone watched "Real Genius", learned what was needed, and pulled a Lazlo Hollyfeld.

[imjonse]

YOLO 9 and YOLO 10 were both published after this paper and according to them they are better than RTDETR.

[wongarsu → imjonse]

And interestingly YOLO 10 also got rid of NMS

https://docs.ultralytics.com/models/yolov10/

[GaggiX]

Just to clarify this paper was released in April 2023.