Multi-instance Reservoir Sampling and Selection for Online Continual Detection over VHR Remote Sensing Images

Abstract

Convolutional neural networks (CNNs) have shown outstanding performance in object detection over very-high-resolution (VHR) remote sensing images. However, the regular offline learning mode suffers from catastrophic forgetting problems and performs poorly on the non-stationary and never-ending data. To address this issue, a multi-instance reservoir sampling and selection method (MIRSS) is proposed for the continual detection on continuously generated remote sensing images. A multi-instance reservoir sampling module is used to build a size-fixed buffer and stores the previously learned samples for memory consolidation. Meanwhile, the situation that several objects may exist in each class of an image is focused. Moreover, samples in the buffer are selected with the reservoir selection module for retraining detectors. The experimental results based on three publicly available VHR satellite images, including images from the NWPU VHR-10, RSOD and DOTA data sets, highlight the effectiveness and practicality of the method.