Open AI-ML-Enthusiast opened 3 years ago
I may not be right but still...
the Focal Loss is applied to the MRCNN class loss because it helps in dealing with the class imbalance issue specifically for the object detection and instance segmentation tasks. These tasks involve determining the presence or absence of multiple object classes in an image, and class imbalance is a common challenge in such scenarios.
On the other hand, the other losses such as RPN class loss, RPN bounding box loss, MRCNN bounding box loss, and MRCNN mask loss are not affected by severe class imbalance issues. These losses are more directly related to the localization and segmentation tasks.
Thanks for the nice repo. Maybe you applied the Focal Loss in mrcnn_class_loss_graph, but what about rpn_bbox_loss_graph and rpn_class_loss_graph ? Why did you not apply here?
Thanks