What the T stands for in TIou?

[plemeri]

What the T stands for in TIou?

[sanghyun0927]

chatGPT says "In the context of the provided code, the 't' in TIoU stands for "thin," as the TIoU class calculates a thin (boundary-based) version of the Intersection over Union metric, similar to the BIoU class.". Think he's right.

[plemeri]

Hi, TIoU was something I tried to make my own boundary related evaluation metric which I failed.
The evaluation script in this repository is identical to the BIoU.

Sorry for the misleading content 😢

[sanghyun0927]

Oh, my.. got it!

[sanghyun0927]

I am becoming more inclined to trust avgTIou over mBA in evaluating test scores. I'd like to continue to use avgTIou as an evaluation metric, but I'm curious what the author thinks.

[plemeri]

As I mentioned before, TIoU is identical to the BIoU only for this repository, and I was also worried about considering the right boundary metric. In my opinion, BDE, which was used in PGNet, was not a good option for boundary metric since it wasn't quite stable enough for measuring the performance and there wasn't any accessible official evaluation script. On the other hand, BIoU and mBA were producing quite stable values and we could easily find official evaluation script, so I chose one of them.

Long story short, I think BIoU and mBA are both okay to use. Maybe BIoU is more appropriate measure for pixel-wise prediction since they are both identical in terms of measuring only around the boundary area since BIoU measures IoU around the edges (which is more common in this field) while mBA measures accuracy only.

Thanks for sharing your thought.