How does SAHI algorithm accurately merge overlapping detection boxes from different slices?

[Mario-Ancx]

Hello community,

I'm currently working with the SAHI library for object detection on large-scale or high-resolution images. I understand that SAHI uses intelligent algorithms to ensure accurate merging of overlapping detection boxes from different slices, which maintains the integrity and precision of the detected objects.

However, I would like to know more about how exactly this algorithm works. Specifically, how does the algorithm determine that overlapping detection boxes from different slices belong to the same object rather than two separate objects that are close to each other in position?

If anyone could provide insights into this or point me to the specific part of the code where this algorithm is implemented, I would greatly appreciate it. Thank you!

Best regards