Why we can change input image size of darknet YOLO after training?

Question

Darknet yolo configuration files is something like this:

[net]
batch=64
subdivisions=8
height=416
width=416
channels=3
...

But we can change height and width without training the model and get a better result with higher image size (of course inference takes longer time). I'm a little confusing. What happens for weights when we change image size? For example if the first layer has N inputs and the second layer has M neurons then we should have (N*M + 1) weights, if the network is fully connected. So the question is: how can we change N (width and height) without changing the weights?

Answer 1

I'm also rather new to the topic, but I think that only Yolov1 uses fully connected layers. Have a look at the architecture of e.g. Yolov2. There are no FC layers, only convolution. This might be the answer?