Paper writing, documentation update and RetinaNet reading
Week dedicated to finish writing the paper and reading more papers. The paper in general was updated, reviewing each section and updating figures. The documentation was also reviewed.
RetinaNet
RetinaNet [1] paper is a new approach in object detection. It discusses the problems of one-step approaches and proposes improvements that are reflected in RetinaNet. A new loss function is proposed and class imbalance during training is identified as one of the main problems of one-stage detectors. This means that they evaluate 10^4-10^5 candidate locations but the amount of objects is low. The new loss function pays less importance to easy objects, those that are not difficult to be found and classify correctly.