Oriented Object Detection in Aerial Images Based on the Scaled Smooth L1 Loss Function

Abstract

Although many state-of-the-art object detectors have been developed, detecting small and densely packed objects with complicated orientations in remote sensing aerial images remains challenging. For object detection in remote sensing aerial images, different scales, sizes, appearances, and orientations of objects from different categories could most likely enlarge the variance in the detection error. Undoubtedly, the variance in the detection error should have a non-negligible impact on the detection performance. Motivated by the above consideration, in this paper, we tackled this issue, so that we could improve the detection performance and reduce the impact of this variance on the detection performance as much as possible. By proposing a scaled smooth L1 loss function, we developed a new two-stage object detector for remote sensing aerial images, named Faster R-CNN-NeXt with RoI-Transformer. The proposed scaled smooth L1 loss function is used for bounding box regression and makes regression invariant to scale. This property ensures that the bounding box regression is more reliable in detecting small and densely packed objects with complicated orientations and backgrounds, leading to improved detection performance. To learn rotated bounding boxes and produce more accurate object locations, a RoI-Transformer module is employed. This is necessary because horizontal bounding boxes are inadequate for aerial image detection. The ResNeXt backbone is also adopted for the proposed object detector. Experimental results on two popular datasets, DOTA and HRSC2016, show that the variance in the detection error significantly affects detection performance. The proposed object detector is effective and robust, with the optimal scale factor for the scaled smooth L1 loss function being around 2.0. Compared to other promising two-stage oriented methods, our method achieves a mAP of 70.82 on DOTA, with an improvement of at least 1.26 and up to 16.49. On HRSC2016, our method achieves an mAP of 87.1, with an improvement of at least 0.9 and up to 1.4.