Skip to main content
. 2024 Feb 24;14:4489. doi: 10.1038/s41598-024-54993-y

Figure 6.

Figure 6

Predicted class weights from the teacher network for two IPB configurations, as well as the actual training dataset distribution. The multi-scale input patches from multi-label images (ML+MS) allows the teacher to learn a more accurate distribution of the image classes, which it can then use to provide the student with a more accurate pseudo-balanced dataset.