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

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