Dear research community?
How are the anchor sizes defined in the cfg files of YOLOv3 and YOLOv4 object detectors? For example from the Yolov4 cfg file
anchors = 12, 16, 19, 36, 40, 28, 36, 75, 76, 55, 72, 146, 142, 110, 192, 243, 459, 401
During training time do these value get overridden somehow? to the best of my knowledge, the main CNN code simply read these numbers and used them for Non Max suppression.
Any insight to the details of defining the anchors would be greatly appreciated.
With best,
Azmyin
I believe the following resource might be a good starting point to understand how Anchor boxes are generated.
1. https://d2l.ai/chapter_computer-vision/anchor.html
2. https://towardsdatascience.com/anchor-boxes-the-key-to-quality-object-detection-ddf9d612d4f9
3. https://www.youtube.com/watch?v=vRqSO6RsptU&ab_channel=ValentynSichkar (from 15:09 timestep) ***
- Badges
- Science topic
- Similar topics
- Filing