yes, by the way am trying to add image. how can i do that. am failing

I don't really get what you need? We are dealing with segmentation not object detection

Theres an idea am trying to implement

Is it not that we are supposed to get the segmentation of the green area???

If I get you right, you want the bounding boxes first before implementing segmentation?

Yes

No some have the brown, green, yellow and at times blue.

I already Know wrote an algorithm that draw that bounding box above

My confusion is whether the bounding the border edges is affecting IoU

I think most of us did not train an object detection model for that but instead used a segmentation model.

Did not train an object detection model, its an algorithm

Yes it is definitely. Mask it and see the difference in IOU per ground truth.

Let me try working something out

Thanks for responding

GoodDay

You are welcome brother!

No it is not going to affect the IoU but if you want to be safe add some padding to the box?

Bro it does, From what he wants to do. He uses the bounding boxes to get the exact polygon boundaries before segmenting so that it looks like the masked image of the ground truth - If he pads or he doesn't pad, it will still distort the shape of the polygons do you get it? It will not distort it, if the algorithm gets the exact bounding box. If not, it will distort it by a 4% - God knows how much😂.

So that he will use the ground truth given to train? 😂, did not think that is what he was trying to do, no way it's working then, Distortion in some of the shapes will be fatal to IoU

Wait, it seems his algorithm captures the most extreme ends of the polygon, he can convert the ground truth to make the mask border on the extreme edges and he will get a perfect overlap, if his algorithm is accurate then it will

Noted with thanks,

Now compare it to the ground truth, if it is above 97 IOU, you are likely doing it right

@offei_lad that is very true!