search

AlexeyAB / darknet

YOLOv4 / Scaled-YOLOv4 / YOLO - Neural Networks for Object Detection (Windows and Linux version of Darknet )
http://pjreddie.com/darknet/
Other
21.8k stars 7.97k forks source link

How to improve recall when there are too many objects in a row? #1123

Open zhliu2018 opened 6 years ago

zhliu2018 commented 6 years ago

I'm doing number detection,when there are many objects in a row ,such as 12 ,it's often missing some object.But it do well when there are less than 8 numbers in a row.Here are some picture may help you understand.I am using yolo2,and my cfg file is as follow: height=288 width=416

[region] anchors = 1.08,1.19, 3.42,4.41, 6.63,11.38, 9.42,5.11, 16.62,10.52 bias_match=1 classes=10 coords=4 num=5

image image Any suggestions will be appreciate!

AlexeyAB commented 6 years ago
zhliu2018 commented 6 years ago

Thanks for the reply. @AlexeyAB

recalculate anchors by using this command:

I have recalculated anchors by using this command: darknet.exe detector calc_anchors data/obj.data -num_of_clusters 5 -width 416 -height 288 and divede it by 32.

then set thresh = .6 in the [region]-layer and train again from the beginning

I am using thresh=.6 now, and I want to know this value of thresh is for probability or NMS?

What mAP can you get currently?

for thresh=0.25, mAP=85%, recall=88%

Attach your cfg-file

yolo-obj.zip

finally I think this problem may caused by the feature map , because the feature map size is 13*7 after many downsample, and there are as many as 12 objects in a row. Could you give me some advice on how to improve the size of feature map except increasing the size of input image.

AlexeyAB commented 6 years ago

I have recalculated anchors by using this command: darknet.exe detector calc_anchors data/obj.data -num_of_clusters 5 -width 416 -height 288 and divede it by 32.

Can you show screenshot of result of this command? darknet.exe detector calc_anchors data/obj.data -num_of_clusters 5 -width 13 -height 9

Also try to use int the [region] layer and train about 6000 iterations:

thresh = .7
random=1

I think this problem may caused by the feature map , because the feature map size is 13*7 after many downsample, and there are as many as 12 objects in a row. Could you give me some advice on how to improve the size of feature map except increasing the size of input image.

Just increase width=608 in your cfg-file.