search

NVIDIA / caffe

Caffe: a fast open framework for deep learning.
http://caffe.berkeleyvision.org/
Other
672 stars 263 forks source link

clustering-FOV-bug-fix #597

Open MichalZajacMTS opened 1 year ago

MichalZajacMTS commented 1 year ago

receptive field of view for clustering was not correctly updating according to input image size - fixed

MichalZajacMTS commented 1 year ago

Hi, I've had problems with detection on images with greater height than training images, and images with greater height then width. For example here, I have trained the net on images with halve the height than the tested one and the net was able to find only the top row objects. image

After applying my change, I can work with any image size with no problems: image

I have been using this patch for few years now, so it's tested on many cases. Nobody else seems to have noticed this so maybe I'm doing something wrong - in this case please guide me to find my mistake. Thank you.