Accepted article at MDPI Remote Sensing
Correcting Imprecise Object Locations for Training Object Detectors in Remote Sensing Applications
06.12.2021
Authors
Maximilian Bernhard, Matthias Schubert
Abstract
GPS annotations are often not well aligned with georectified imagery and also not in the format of bounding boxes, posing an obstacle for training object detectors. We propose a correction framework that allows to robustly learn class activaton maps from such noisy point labels and refine the annotations. When applied as a preprocessing step on the raw annotations, we can greatly improve the performance of object detectors. [more]
