0
$\begingroup$

Weight Cross-Entroy (WCE) helps to handle an imbalanced dataset, and Cityscapes is quite imbalanced as seen below:

enter image description here

If we check the best benchmarks on this dataset, most of the works use bare CE as a loss function. I don't get it if there are any special causes that would lead WCE to a worse result for semantic segmentation tasks on the mIoU evaluation.

I'm especially asking because I'm working in an even higher unbalanced dataset (multi-minority classes on the ratio of 1:1000 to the majority classes) and got very surprised when bare CE outperformed WCE on the mIoU metric.

I found so far that WCE can yield many false positives from minority classes, but beyond that, would there be more reasons for it?

Improve this question
$\endgroup$
5
  • $\begingroup$ What issue do you see regular crossentropy having when it comes to imbalanced classes? $\endgroup$ Commented Jun 1, 2022 at 20:57
  • $\begingroup$ In a highly imbalanced class situation, wouldn't the learning be tendentious to predict only the most dominant classes or generalize badly the minority classes? I mean, that's not why many works use WCE? $\endgroup$ Commented Jun 1, 2022 at 21:07
  • $\begingroup$ In a highly imbalanced class situation, aren't the dominant classes more likely? $\endgroup$ Commented Jun 1, 2022 at 21:09
  • $\begingroup$ Do you have any reference or explanation that CE is not an issue in this scenario? B/c my doubt is that I used to see that regular CE was a problem on this, and that's why some works suggested WCE (like U-net paper); Dice, and IoU surrogate losses. $\endgroup$ Commented Jun 1, 2022 at 21:15
  • 1
    $\begingroup$ To get started with probability prediction and how that interacts with class imbalance, I suggest starting with two blog posts by Frank Harrell, founding chair of the Department of Biostatistics and Vanderbilt Medical School: Post 1 Post 2. These will not directly answer your question, but they will start you down the rabbit hole. $\endgroup$ Commented Jun 1, 2022 at 21:34

0

Reset to default

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Start asking to get answers

Find the answer to your question by asking.

Ask question

Explore related questions

See similar questions with these tags.