search

wangyuxin87 / ContourNet

A PyTorch implementation of "ContourNet: Taking a Further Step toward Accurate Arbitrary-shaped Scene Text Detection" (CVPR2020)
Other
227 stars 46 forks source link

About alphashape algorithm and NMS #11

Closed DREAMXFAR closed 4 years ago

DREAMXFAR commented 4 years ago

Hello, I have read your paper and I think it is a quite good job. However, I met some problems when doing NMS and alphashape. As far as I know, alphashape algorithm relys on the setting of alpha and the density of the points. What's more, the number of points is not steable after NMS. I finds that I sometimes get a better result if I don't use NMS.

So how do you find suitable parameters in Alphashape algorithm? And I want to know why NMS is necessarry, will NMS bring some problems?

Thank you very much~

wangyuxin87 commented 4 years ago

Thanks for your attention. It seems that you meet this problem when using Alpha-Shape method to represent curved text. We firstly set a large value to the parameter in Alpha-Shape to guarantee a more complete region representation and then use some tricks to refine the region based on the contour points. In addition, inspired by some key-point detection tasks, we use NMS to find the extreme points in the contour region and we find it works well in our experiments. I hope this can help you.

DREAMXFAR commented 4 years ago

Thank you for your answer. I am still very interested in how to refine the region. I will appreciated it if you can give some further details about the tricks you used. Do you find a different alpha for every image, or just try to find an alpha that suitable for all test images?

wangyuxin87 commented 4 years ago

I am sorry that we do not plan to released these tricks. We just set a fixed alpha for all images.