Mitigating the effects of severe imbalance in multi-class semantic segmentation

Class imbalance is one of the main weaknesses in modern machine learning methods.
In this area, datasets with an imbalance ratio greater than 1:100 are defined
as severely imbalanced. These require specific precautions and techniques to deal
with the issue.
In this thesis, different approaches to tackle the problem of severely imbalanced
datasets in semantic segmentation are explored. Solutions such as resampling,
the One-vs-Rest approach, and loss change are implemented and compared
discussing their benefits and drawbacks. Furthermore, the delicate evaluation process
is explained in all its complexity giving specific weight to the obtained results.