AlexeyAB
commented
4 years ago Why did you keep the anchors in yolov4? Could elaborate a little bit on this configuration.
Because it achieves higher accuracy and FPS. We did not conduct a full-scale comparison of all possible anchor-free and anchor-based solutions.
-
anchor-free - is good, when you just can't use anchors. Or f.e. you can use anchors for bboxes and anchor-free approach for keypoints/landmakrs: https://github.com/ouyanghuiyu/darknet_face_with_landmark
-
anchor-based - is good when you can use it, and then you can adjust other points/segmentation inside bbox like YOLACT++: https://github.com/dbolya/yolact and https://arxiv.org/pdf/1912.06218.pdf
As you talked about in the Yolov4 paper anchor-free methods seem a good technic.
Why do you think this is a good technic?
As you talked about in the Yolov4 paper anchor-free methods seem a good technic. Why did you keep the anchors in yolov4? Could elaborate a little bit on this configuration.