RAISE: Rank-Aware Incremental Learning for Remote Sensing Object Detection

The deep learning method is widely used in remote sensing object detection on the premise that training data have complete features. However, when with a fixed class are added continuously, trained detector less able to adapt new instances, impelling it carry out incremental (IL). IL has two tasks knowledge-related symmetry: continuing learn unknown knowledge and maintaining existing knowledge. Unknown more likely exist these which features dissimilar from those of old instances cannot be well adapted by before IL. Discarding all leads catastrophic forgetting knowledge, can alleviated relearning while different subsets represent ranges memory-retention effects Due values data, methods without appropriate distinguishing treatment preclude efficient absorption useful Therefore, rank-aware instance-incremental (RAIIL) proposed this article, pays attention difference aspects data-learning order loss weight. Specifically, RAIIL first designs rank-score according inference results true labels determine then weights balance contribution. Comparative analytical experiments conducted public datasets for detection, DOTA DIOR, verified superiority effectiveness method.