YOLOv4 image resizing

[ghost]

I am trying to understand the concept of resizing input images to fit network size. I have been through several issues in this Github repository and I haven't seen a clear explanation.

I, let's say, have an image of 1920 x 2000 and the network size is 416 x 416, does that mean the input image will be "compressed/squeezed" into a smaller shape. Wouldn't that mean that small objects might be squeezed into almost 1 pixel ?

If I can have the link to the code sample where the squeezing is performed to further understand this concept, that would be appreciated.

Thank you !

[stephanecharette]

See: https://www.ccoderun.ca/programming/darknet_faq/#square_network and https://www.ccoderun.ca/programming/darknet_faq/#time_to_train

If your objects become "almost 1 pixel" when you resize to your network dimensions, then Darknet/YOLO wont find them.

[ghost]

See: https://www.ccoderun.ca/programming/darknet_faq/#square_network and https://www.ccoderun.ca/programming/darknet_faq/#time_to_train

If your objects become "almost 1 pixel" when you resize to your network dimensions, then Darknet/YOLO wont find them.

Thank you ! This is answers a lot of questions I had 👍🏼

[rabsym]

See: https://www.ccoderun.ca/programming/darknet_faq/#square_network and

Before reading this I thought yolo keeps aspect ratio, ie. image 1280x720 will be resized to 416x234, and then inserted into 416x416 network.
Maybe I was wrong :?

[stephanecharette]

You are wrong. As described in the link, it does not maintain aspect ratio.

[TobiasVorsmann]

If you set letter_box = 1 in [net] section it keeps the aspect ratio (section For training only -> Data augmentation -> letter_box: https://github.com/AlexeyAB/darknet/wiki/CFG-Parameters-in-the-%5Bnet%5D-section)

[lsd1994]

Here is the difference between resize and letter_box.
#232 (comment)