search

georgesung / ssd_tensorflow_traffic_sign_detection

Implementation of Single Shot MultiBox Detector in TensorFlow, to detect and classify traffic signs
MIT License
530 stars 222 forks source link

resized images or full images #10

Closed hbbeytur closed 7 years ago

hbbeytur commented 7 years ago

Thank you for your work. For several days I'm trying to understand and implement your work. I read the paper and I get confused. Are you using the convolution type scanning on the resized image or on the original image. When I looked at the paper it seems to be the original images but in the code I think you are using the resized images in data_prep.py. The problem come through when the calc_iou() tries to find the matched, but the rescaled box coordinates of the ground truth aren't in between 0 and 1. Can you clarify this issue a little?

Thanks

hbbeytur commented 7 years ago

or should we crop the signs from original images and resize them?

hbbeytur commented 7 years ago

I understood the problem. Mistakenly I think that the signs should be cropped and resized but the resized images are the original images