I am trying to train a YOLO model to draw bounding boxes around individual handwritten words. I don't need the network to classify them and I don't need the words to be in order either. I would have to use a fine grid (maybe 20x35) to detect every word on a piece of lined paper. I am planning on inputting 600x800x1 pixel images. I found the image below of a simple architecture designed to work with only one class. I just need some help tweaking it so it works with my input image and grid size. So far I've added a 16x7x7-s2-p3 layer to the top to get a 300x400x16 image, and then proceeded with the rest like normal. This means the final output before the fully connected layers is 4x6x30. My main question is, does the final output before the fully connected layers need to match the grid size? If so, how can I change the structure to get it to match. Does it matter if the amount of inputs into the fully connected layers are less than the amount of outputs (4x6x30=720 inputs when flattened and I have 20x35x5=3500 outputs). Any advice would be greatly appreciated.