Sharing NMM ideas for more efficient results #831
YoungjaeDev
started this conversation in
Ideas
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
-
Hi!
First, sahi brought great innovation to the Tiling Inference.
It has proven to work reliably in air view images, mostly brid views.
In most cases(like bird view), there is no occlusion mostly.
However, in the case of images viewed from the front, the case of occlusion is often seen
So when I go back and think about the purpose of doing NMM, the purpose of NMM is to "merge on the part that's cut by Slicing."
So for example, if there are 9 slices, I think slice 1 should be NMM only with 2, 3, 4, 5, 6, 7, 8, 9
This premise, of course, is that the forward->nms result (From yolov5, Autoshape) for slice 1 is perfect, And I don't think there's anything to touch here either. Only with the other boxes. Therefore, I don't think there is a need to proceed with box results and NMM within slice 1. For example, when there is occlusion, there is a case where the bbox between the slice 1 result is merged, which seems to be a problem.
What I've said seems to be an algorithm that can be applied to both the air view and the front view. What do you think?
Beta Was this translation helpful? Give feedback.
All reactions