Closed xsvonjhs closed 5 years ago
@xsvonjhs I think the problem is because a single digit is very short, so it should be captured by the square anchors. But the smallest square anchors is 30x30 pixels, maybe your single digit is too small to be captured by the anchors. There are two solutions: (1) increase the input image size in inference. (2) In the first 1/8 feature map, add additional small anchors to capture the small short texts.
@Yuanhang8605 thanks for your reply! i think the 2nd method sounds practical to me, but i'm not sure where is the code, any hints?
@xsvonjhs in models/deploy.prototxt min_size is 30 change that to a smaller value
Hi,
I am using demo.py to run detection on my own images, but it seems like single digit is ignored. I understand that it is somehow being treated as noise, but it is quite important in my case. Is there any workaround?