Use of gap in ultralytics.data.split_dota.get_windows

[dimidagd]

What would be the use of gap in ultralytics.data.split_dota.get_windows ? Why would one not want a negative gap (overlap) between crops?

ultralytics/ultralytics/data/split_dota.py

Line 89 in 61088b4

def get_windows(im_size, crop_sizes=[1024], gaps=[200], im_rate_thr=0.6, eps=0.01):

[glenn-jocher]

@dimidagd the gap parameter in ultralytics.data.split_dota.get_windows defines the distance between adjacent sliding window crops. A positive gap means there is a space between crops, while a negative gap (overlap) ensures that adjacent crops share some common area. Overlapping can be beneficial as it provides more context around the edges of each crop, which can help the model make more accurate predictions, especially for objects that might be cut off at the edges of a crop. Here's a quick example of how you might use a negative gap:

from ultralytics.data.split_dota import get_windows

# Assuming you have a 'size' and 'image_shape' defined
windows = get_windows(size=size, image_shape=image_shape, gap=-50)

In this snippet, each window will overlap the next by 50 pixels. Hope this helps! 😊👍

[dimidagd]

Cool, thanks. Any reason why the default value is positive?

[glenn-jocher]

Hi there! The default positive gap value is set to avoid redundancy and speed up the inference process by reducing the number of crops. Overlapping windows (negative gap) can increase accuracy for objects on the edges but also increase computation. It's a balance between performance and speed. You can always adjust it based on your specific needs. Happy coding! 😊🚀