Image input size

[kamadforge]

What is the input size to Faster R-CNN framework? The paper says the image is resized so that the smaller side is 600 pixels long. But the network has to take images of fixed size, doesn't it? And in the test time the images are 224x224. So the detection is run on smaller images?

[chankim]

All the layers' input /output (bottom/top) blob sizes are determined from first image blob size. Before processing the forward step, it reshapes for the input blob size. If the original input image is too big, it is resized down so that smaller side is 600 pixesl long but if it is too small it is expanded if I remember correctly.