search

fizyr / keras-retinanet

Keras implementation of RetinaNet object detection.
Apache License 2.0
4.38k stars 1.96k forks source link

Impact on bounding boxes when resizing the images for training #1515

Closed hlmhlr closed 3 years ago

hlmhlr commented 3 years ago

Hi,

I have two different sizes of images in my dataset (960x540 1360x765) which are resized as image_min_side = 800 and image_max_size = 1333 before training.

I am puzzled that, would this resizing of images before training have not any impact on objects bounding boxes (x,y, w, h)? If resizing of bounding boxes is also done in the retinanet code, could someone please point out that portion?

Since, i did an experiment to resize the images individually without bounidng boxes but what i observed, bounding boxes and resized image were totally shifted when i plotted them together.

i would be very thankful for clarifying this point.

thanks,

hlmhlr commented 3 years ago

Hi,

I have two different sizes of images in my dataset (960x540 1360x765) which are resized as image_min_side = 800 and image_max_size = 1333 before training.

I am puzzled that, would this resizing of images before training have not any impact on objects bounding boxes (x,y, w, h)? If resizing of bounding boxes is also done in the retinanet code, could someone please point out that portion?

Since, i did an experiment to resize the images individually without bounidng boxes but what i observed, bounding boxes and resized image were totally shifted when i plotted them together.

i would be very thankful for clarifying this point.

thanks,

Hi,

Resizing of images would have direct impact on bounding boxes and bboxes should also be resized. In fizyr/keras-retinanet, bboxes are also resized before training.

The portion of code where bboxes are resizing is mentioned in the issue. I explored before posting this question but couldn't get the answer. Well, its resolved now and i am closing this issue as well.

Thanks,