search

YonghaoHe / LFFD-A-Light-and-Fast-Face-Detector-for-Edge-Devices

A light and fast one class detection framework for edge devices. We provide face detector, head detector, pedestrian detector, vehicle detector......
MIT License
1.31k stars 330 forks source link

license plate accuracy of v1_small #81

Open wang-xinyu opened 4 years ago

wang-xinyu commented 4 years ago

@YonghaoHe thanks for your work.

I meet a strange problem in license-plate detection, for the same image, when resize it to 320x240, the confidence of the box is 0.99+, but when resize it to 640x480, the confidence only 0.3+. Do you have an idea why this can happen?

WechatIMG501

regards

YonghaoHe commented 4 years ago

@wang-xinyu if the target license plate is in the range [64, 512], it should be confidently detected. You can resize the image to some other scales, and observe the results. If the problem remains, I guess the failure may be caused by different data distribution. Note that the CCPD dataset is collected in Anhui province.

wang-xinyu commented 4 years ago

@YonghaoHe Thanks a lot.

For the range[64, 512], does it mean the plate width in the image?

For example:

YonghaoHe commented 4 years ago

@wang-xinyu I means the longer side, in license plate detection, it should always be the width. In your description, no matter in 640x480 or 320x240, the plate should be well detected. I think the train data matters. You can try some other images to see if the same problem occurs. If you have your own data, you can finetune a new model.

wang-xinyu commented 4 years ago

@YonghaoHe Yes, I understand, thanks a lot!