Copy-paste Augmentation for better saliency learning of underrepresented Object Classes
Zafar Aziz, HyungWon Kim · 2023
Data Augmentation plays an important role in training a better and well generalized object detection model. The publicly available object detection datasets are limited and are severely unbalanced. This leads to an inconsistent and unstable representation learning of the minority class objects. Hence, the resulting object detector model suffers from confirmation bias problem, giving a lot of inconsistent detections of minority class samples. To solve the unbalanced nature of object detection datasets we propose copy-paste augmentation method. In this method we paste the minority class samples into the images, which have no minority class samples. Our method relies on the random resizing and brightness adjustment of the selected minority sample for pasting. Our method can increase the number of samples of minority class which in turn help detector to learn better and accurate representation of the minority class objects.