search

microsoft / CNTK

Microsoft Cognitive Toolkit (CNTK), an open source deep-learning toolkit
https://docs.microsoft.com/cognitive-toolkit/
Other
17.53k stars 4.28k forks source link

Wrong Bounding Boxes with smaller sized Faster R-CNN #3581

Open jkooper opened 5 years ago

jkooper commented 5 years ago

Dear CNTK Community,

I am trying to train Faster R-CNN as explained here: https://docs.microsoft.com/en-us/cognitive-toolkit/object-detection-using-faster-r-cnn#run-faster-r-cnn-on-your-own-data Basically all works and evaluates fine when using the default __C.IMAGE_WIDTH = 850 __C.IMAGE_HEIGHT = 850

But I want to train with 416x416 to better compare to YOLO. When setting width and height to 416x416 and train with 1920x1080 images and test e.g. with 1024x1024, the output gets scaled to 800x800 and squished falsy bounding boxes appear as shown in the screenshot. 0_regr_apple braeburn 1

As said, with the defaults it is fine. Do I do something wrong?

eyagarci commented 5 years ago

@jkooper Can you find a solution? I have the same problem.

jkooper commented 5 years ago

@eyagarci Unfortunately not. I just left the resolution of the network as it was by default. Sorry I can't help.