An Improved Aggregated-Mosaic Method for the Sparse Object Detection of Remote Sensing Imagery

Object detection based on remote sensing imagery has become increasingly popular over the past few years. Unlike natural images taken by humans or surveillance cameras, the scale of remote sensing images is large, which requires the training and inference procedure to be on a cutting image. However, objects appearing in remote sensing imagery are often sparsely distributed and the labels for each class are imbalanced. This results in unstable training and inference. In this paper, we analyze the training characteristics of the remote sensing images and propose the fusion of the aggregated-mosaic training method, with the assigned-stitch augmentation and auto-target-duplication. In particular, based on the ground truth and mosaic image size, the assigned-stitch augmentation enhances each training sample with an appropriate account of objects, facilitating the smooth training procedure. Hard to detect objects, or those in classes with rare samples, are randomly selected and duplicated by the auto-target-duplication, which solves the sample imbalance or classes with insufficient results. Thus, the training process is able to focus on weak classes. We employ VEDAI and NWPU VHR-10, remote sensing datasets with sparse objects, to verify the proposed method. The YOLOv5 adopts the Mosaic as the augmentation method and is one of state-of-the-art detectors, so we choose Mosaic (YOLOv5) as the baseline. Results demonstrate that our method outperforms Mosaic (YOLOv5) by 2.72% and 5.44% on 512 x 512 and 1024 x 1024 resolution imagery, respectively. Moreover, the proposed method outperforms Mosaic (YOLOv5) by 5.48% under the NWPU VHR-10 dataset.