I am training an image segmentation model, but I would like my IoU to be improve (maximized) during the training. How can I achieve that?
Please note, I am actually looking for a python code recommendation, where the code helps to implement the Jaccard / IoU as a loss function?
For semantic segmentation of 2D image data, the Intersection over Union (IOU) measure or Jaccard index is usually a performance measure for a segmenter that has been trained using cross entropy or other conventional loss function that is differentiable (i.e. optimisable by stochastic gradient descent) with respect to the parameters of your assumed neural (convolutional) network. For IOU this may be problematic as the measure may be indeterminate.
There is some previous work in this area, e.g. "UnitBox: An Advanced Object Detection Network"
https://arxiv.org/pdf/1608.01471.pdf
My main comment is that IoU is a single-image based pixel-counting measure. By this, I mean that it does not account for temporal and spatial differences. If these aspects are not important in your segmentation problem, then you might get an improvement by attempting to directly train via IoU. At any rate, you should always use a conventional (cross-entropy) loss function for a baseline.
Please recommend this answer if you find it useful, and/or ackowledge if you use it in a publication.
Dear Kenan Morani
This tutorial might also be helpful, have a look
https://www.youtube.com/watch?v=Gc4BAJpjJKA
Kind Regards
Qamar Ul Islam
- Badges
- Science topic
- Similar topics
- Organizational Psychology
- Training