- This paper addresses the issue of neural networks learning from an unbalanced dataset, that mostly seen in single stage object detectors, resulting in their poor performance.
- The Focal Loss , introduced in the paper, is the main contribution of the paper, that improves the performance of single stage object detectors that dynamically scales the cross entropy loss to compensate for the imbalance.
- Most approaches resolve this issue by implementing OHEM, that forces the network to investigate false postives and false negative examples, ie, FG samples for which the the network predicited a BG with high probability and BG samples for which the network predicted a FG with high probability.