FurkanOM / tf-faster-rcnn

Tensorflow 2 Faster-RCNN implementation from scratch supporting to the batch processing with MobileNetV2 and VGG16 backbones
Apache License 2.0
94 stars 61 forks source link

Anchors and rpn module for varied image sizes (width!=height) #4

Closed MLForNerds closed 4 years ago

MLForNerds commented 4 years ago

Hi FurkanOM, Thanks for the great work!! This repository works only for square images. What if i want to keep the aspect ratio of my images and train, like 1280x720. Then the anchor generation and anchor target generations have to be modified. Any thoughts on implementing this soon?

FurkanOM commented 4 years ago

Hi, In fact, the anchor generation processes were working the way you specified at first, but I changed it later. You can find the working version that way by examining the previous commits but I don't plan to implement that way. Since the bounding boxes are in a relative format, the size or aspect ratios of the images do not pose any problem.

MLForNerds commented 4 years ago

Hi, Thanks for the response. Can you point me the exact commit in which it is working that way. I can modify my code accordingly. Thanks.

FurkanOM commented 4 years ago

Hi, sorry but i couldn't. I'm dealing with other things right now. You need to find that by yourself.