Dependency of image size on detect with FasterRCNN

I trained a fasterRCNN to detect left and right knees in bilateral knee x-rays using VGG-16. The accuracy reported in step 4 was near 100% at the end of training. The trained network works great - if I resize the input x-ray to just the right size - by trial&error. Is anyone aware of any documentation regarding this dependency and what was done with the image size during training such that it could achieve such high accuracy on the minibatches?