Re-annotation of training samples for robust maritime object detection

Bolstering Maritime Object Detection with Synthetic Data

“…In our previous work [1], [2], [3], we proposed an object detection network tasked with robust detection of two coarse classes, buoys and ships; given an image, a detection consisted of an object bounding box and class confidence. This work extends our efforts of creating a more reliable and robust object detection system [4], [5], by focusing on producing higher quality classification outputs, that is a more precise label (e.g. from boat to sailboat or motorboat) and providing a usable uncertainty metric for said classification.…”