Enhancing class-incremental object detection in remote sensing through instance-aware distillation

Hangtao Feng, Lu Zhang, Xu Yang, Zhi-Yong Liu

Abstract: Highlights•We propose an instance-aware knowledge distillation approach for class-incremental object detection, utilizing a previous model as a teacher to guide learning on new data for incremental detection of new classes without forgetting old classes.•Instance-aware distillation preserves old class responses while aiding new class learning, reducing catastrophic forgetting and enhancing detection in complex backgrounds.•Augmented pseudo-label module expands training instances of old classes, generating pseudo-labels on original and augmented images, reducing forgetting with an IoU-based filter, providing more instances to mitigate catastrophic forgetting.•Experiments on the challenging DOTA, DIOR, RTDOD and PASCAL VOC dataset show our method can incrementally learn new classes while retaining detection accuracy on old classes, effectively alleviating catastrophic forgetting.