whats is NMM and IOS?

[fcakyon]

In the paper we only utilized standard NMS and IOU. The NMM and IOS will be discussed in a follow-up paper.

For now you can consider NMM as non-maximum merging, in other words, it merges the lower scored box instead of suppressing it as in NMS. It is useful when the model confidence threshold is large (as 0.3 to 0.5) and the predictions are confident. It merges the box predictions, corresponding to the same instance, coming from different slices. Note that NMM used in sahi is different from OverFeat in the sense that, NMM merges the boxes by taking their union while OverFeat merges the boxes by averaging them.

IOS is basically intersection over smaller region, ie intersection area of two box / area of smaller box. It is used as the match criteria while performing NMS and NMM.

All NMM and NMS implementations can be performed via IOU or IOS match criteria by efficient matrix calculations using the sahi.postprocess module.

If you have any question regarding the postprocess module of sahi feel free to ask them under this post.

[Grenaf]

Looking forward to the effect of NMM and IOS about smaller region in your new paper，your first paper is very interesting！

[Mo4py]

Do the postprocessing algorithms (NMS, NMM etc.) work based on the bounding box (bbox) or the segmentation mask of each detected mask?

After looking into the code i think it is only based on the bbox coordinates?
Would it not be better to use the mask instead? (So e.g. the IOU is calculated from the intersection of the masks, not the bboxes of each detected instance)

Thanks for every answer and for the great library.