如何修改锚框大小，麻烦问下作者锚框大小和原YOLOv4一样吗 about yolov4-pytorch HOT 7 OPEN

"ANCHORS":[[(1.25, 1.625), (2.0, 3.75), (4.125, 2.875)],
[(1.875, 3.8125), (3.875, 2.8125), (3.6875, 7.4375)],
[(3.625, 2.8125), (4.875, 6.1875), (11.65625, 10.1875)]]
这个比例是如何计算的？

原始的聚类框除以stride就可以得到以上的ANCHORS。

from yolov4-pytorch.

"ANCHORS":[[(1.25, 1.625), (2.0, 3.75), (4.125, 2.875)],
[(1.875, 3.8125), (3.875, 2.8125), (3.6875, 7.4375)],
[(3.625, 2.8125), (4.875, 6.1875), (11.65625, 10.1875)]]
这个比例是如何计算的？

原始的聚类框除以stride就可以得到以上的ANCHORS。

ok，谢谢川大小姐姐，求加QQ还想问下你是咋样学着自己该网络结构的，感觉改的贼6

from yolov4-pytorch.

"ANCHORS":[[(1.25, 1.625), (2.0, 3.75), (4.125, 2.875)],
[(1.875, 3.8125), (3.875, 2.8125), (3.6875, 7.4375)],
[(3.625, 2.8125), (4.875, 6.1875), (11.65625, 10.1875)]]
这个比例是如何计算的？

原始的聚类框除以stride就可以得到以上的ANCHORS。

是聚类得到的框从小到大排序 依次除以8 16 32不？

from yolov4-pytorch.

@argusswift 请问您为什么要这样处理anchors呢，是想达到什么效果的呢？

from yolov4-pytorch.

@argusswift 请问您为什么要这样处理anchors呢，是想达到什么效果的呢？

from yolov4-pytorch.

YOLOv4-pytorch/model/head/yolo_head.py

Line 46 in 863f35c

pred_wh = (torch.exp(conv_raw_dwdh) * anchors) * stride

stride * anchor = actual anchor size

@argusswift 请问您为什么要这样处理anchors呢，是想达到什么效果的呢？

谢谢，懂了！但是原始anchor除以stride并不精确等于现在的anchor，是不是还做了其他处理呢

from yolov4-pytorch.

YOLOv4-pytorch/model/head/yolo_head.py

Line 46 in 863f35c

pred_wh = (torch.exp(conv_raw_dwdh) * anchors) * stride

stride * anchor = actual anchor size

@argusswift 请问您为什么要这样处理anchors呢，是想达到什么效果的呢？

谢谢，懂了！但是原始anchor除以stride并不精确等于现在的anchor，是不是还做了其他处理呢

I don't know for sure but I tend to calculate the anchor by my self. You can also use darknet anchor size and /8 /16 to get the anchor. K-means script might varies a little, so are the results.
But generally, I don't see anchors size would affect final result much, especially when they don't varies hugely. Anchors are just the starting point of the bbox prediction or say, the reference bbox size. Better anchor just help network converge faster imo.

from yolov4-pytorch.