3
$\begingroup$

I was going through the YOLO Object Detection Paper by Joseph Redmon.

The authors use a grid size of $S = 7$. If I am not wrong, the network architecture has carefully been curated for this specific grid size (Since the output of the final conv layer has dimensions $7 \times 7 \times 1024$).

Thus my question is, is there any clear way to adjust the network architecture for different values of $S$? The authors do not mention anything about this in their paper.

$\endgroup$
1
$\begingroup$

Rule of thumb is to make sure the features are downsized to (7 x 7 x no.of filters) in the final conv layer. You can play around with the number of filters in the other layers.

Note that the max pool layers in the middle are responsible for this feature size reduction from 448x448 to 7x7 so be careful while if you're changing their stride & filter size.

If you're thinking about using a different input image size other than 448x448, I would suggest to use YOLO V2 which is FCN based and can work with variable input image dimension.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.