Hacker News new | ask | show | jobs
by bArray 1458 days ago
https://github.com/meituan/YOLOv6/blob/main/docs/About_namin...

> P.S. We are contacting the authors of YOLO series about the naming of YOLOv6.

You should ask _before_ publishing, not _after_.

They claim it runs faster and is more accurate than YOLOv5, yet requires 3x as much computation (GFLOPs)? Something doesn't add up here.

There is unbelievably little information about the architecture too. Unfortunately it's not in a format I can easily throw the cfg in as visualize it: https://gitlab.com/danbarry16/darknet-visual

This appears to be on purpose to advertise DagsHub: https://dagshub.com/pricing
2 comments
saynay 1458 days ago
The individual ops could be faster, so even though there are more of them, the overall speed is quicker. The authors mention using more 16bit ops, so that might be part of the reason?
link
ur-whale 1458 days ago
> There is unbelievably little information about the architecture too

Fair enough, but the repo at git clone https://dagshub.com/nirbarazida/YOLOv6 seems to contain somewhat standard torch code.

So it's not exactly like the architecture is a secret.

The weights is a different story, of course.

link