search

chail / patch-forensics

Investigating patches for fake image classification
130 stars 18 forks source link

The size of image patch on training and testing #4

Closed chenshen03 closed 3 years ago

chenshen03 commented 3 years ago

Hi, thanks for your work!

How do you pre-process the input images in Patch Classifier during training and testing? Random crop to several patches per image, or use the full image as usual?

chail commented 3 years ago

The size of the patch using the Patch Classifier is determined using the receptive field of the classifiers (as the receptive field size is less than size of the full image). These patches are obtained by running the classifier in a fully convolutional manner over the full image as usual, so the output of the classifier is one prediction per patch of the input image.

chenshen03 commented 3 years ago

Thank you very much!