Responding to the Controversy about YOLOv5
blog.roboflow.ai
blog.roboflow.ai
I incorrectly assumed some relationship between @rocauc and the guys behind YOLOv5 due to some shared language and diagrams, sometimes our brains jump to conclusions - sorry. Unfortunately the edit timed out before I could do much about it.
I really appreciate the follow up and the hard work that has gone in since by RoboFlow. This controversy isn't their fault and they really shouldn't get a bad rep for it - in fact their handling of the situation has been good. Maybe some more due diligence, but we're all guilty sometimes. These are the sorts of actions that earn my respect in terms of being an authoritative source.
My original comment was quite reactionary, the YOLOv5 guys have done some interesting work. We now have two new interesting approaches based on YOLOv3, I look forward to reading more about the YOLOv5 internals in the future. There was some strangeness around comments being deleted from their GitHub issues once they started getting noticed, but I'll give them the benefit of the doubt.
With regards to naming, I really think there needs to be some clear change. YOLOv5 does not build on top of YOLOv4 in any way. I appreciate that the team was "scooped" on the name, but I don't think bumping the version number was the correct approach. It's extremely weird to people coming across this for the first time that YOLOv4 can outperform YOLOv5.
There are still some open issues to be addressed with regards to which network performs better and in which use cases, hopefully in the coming weeks it can be worked out.
re: GitHub comment deletion - We determined we should engage when we have easier to reproduce results, so we moved quickly to share them. That's what this post and diligence is, and I've re-engaged on the issues thread in question [1] with fuller remarks + reproducible Colab notebooks.
[1] https://github.com/AlexeyAB/darknet/issues/5920#issuecomment...
> conversation has evolved to focus on architecture and
> performance rather than naming alone.
Agreed. Maybe it wasn't so clear originally, but part of the naming issue came from the point that a fair comparison hadn't been made with the "previous" model.
> re: GitHub comment deletion
This was referring to the comments in the GitHub issues of Ultralytics. But in any case, benefit of the doubt is given. Some people doubted their association with the project and the comments are now gone - so let's see.
The controversy surrounded around two points:
• Should the model truly be named YOLOv5?
• Are the initial benchmarks Roboflow published accurate and reproducible?
We have published more details surrounding our methodology, clear steps and notebooks for reproducing our results, a correction of an unfair comparison we made (re relative inference speed and batch sizes), how we think about the naming controversy and a request for the community to weigh in.
[1] https://news.ycombinator.com/item?id=23478151
[2] https://github.com/AlexeyAB/darknet/issues/5920#issuecomment...