- Receptive field of the Tiny-Yolo is 446x446 pixels, so if you use tiny-yolo-voc.cfg with default widh=416 height=416 with final grid of activations 13x13, then center cells of the final activations see whole image.
Each of the final activations (13x13) see overlaped areas on the input image.
How to calculate it - you should select one of the final activation (one of 13x13), and move layer by layer to the input, so initial receptive field =1
- each convolution layer with kernel=1 doesn’t change the current receptive field
- each convolution layer with kernel=3 adds 2 to the current receptive field
- each maxpool layer with stride=2 multiple 2 the current receptive field
So you will get this table: