how to deal with occlusion when nmm is applied #613
Replies: 4 comments 1 reply
-
@junghyun-avikus if the classes are same, yes this limitation applies. You can use NMS instead of NMM to handle these situations better: Line 163 in 072ea62 |
Beta Was this translation helpful? Give feedback.
-
Yes, but the thing is that the way I am feeding input image to a network is 3 tiled (divided into slices) images and 1 full image. In this case, NMS won't get rid of a large object that presents across the entire frame. So I was trying to come up with an idea that can solve the issue, but I could't come up with one. BTW, slicing works really well for detecting smaller objects and in general. I appreciate your great work. |
Beta Was this translation helpful? Give feedback.
-
@junghyun-avikus how about using NMS with IOS match metric: https://github.com/obss/sahi/blob/main/sahi/predict.py#L132 |
Beta Was this translation helpful? Give feedback.
-
@fcakyon Yes, but that creates another problem that objects in full frame might get removed when parts of an object in tilled frame have higher confidence score. |
Beta Was this translation helpful? Give feedback.
-
I guess when two objects are occluded (1 small object, 1 large object), nmm will merge those to objects into one bounding box. (In case where the smaller object lies on the other larger object). Are there any ideas or potential solutions to this issue?
Beta Was this translation helpful? Give feedback.
All reactions