Darknet: Rules for setting width and height in yolov3.cfg

[ivomarvan]

I tried to use yolov3-tiny on Jetson Nano. It worked without problems.
But I want a network that is small and more accurate.

I tried YOLOv3-320 (and 240x240) but it was too big for GPU memory.
I also tried my own configuration with parameters:
width = 160
height = 160
That worked well, though of course, the accuracy is smaller.
_But it was more accurate than "yolov3-tiny"._

What all combinations of "width" and "height" are allowed?
What does it depend on?

I'd like to try to find the limit that most resolution will work on Jetson Nano.

Thank you for your answer.

[Liedermaus]

If I remember correctly, width and height must be multiples of 16. But I'm afraid it is not that easy, because the accurcy also depends on the number of networks/planes you use and the degree of freedom (filters) you allow in each plane/network). And there is also a STRONG dependency on the training....

[ivomarvan]

Thank you for your answer, Liedermaus.

I have tried that the magic number is not 16, but 32.
Width and height must be a multiple of 32.
I also tried speed and accuracy for different resolution combinations for original net yolov3 and for yolov3-tiny.

For my purposes, the best is yolov3 network with 192x192 resolution. It is really applicable on Jetson Nano. And it's the most accurate from the available configurations (for this moment of my experiments).

[Liedermaus]

It think this is interesting, but it depends strongly on the application. For a network with 192x192 your objects must be fairly big. For my purpose I noticed, that objects are hard to detect below 40x40 pixels. That would be a big part of your input network, so you are hunting cars and elephants ;-)