HVConv : horizontal and vertical convolution for remote sensing object detection

Abstract

Generally, the interesting objects in aerial images are completely different from objects in nature, and the remote sensing objects in particular tend to be more distinctive in aspect ratio. The existing convolutional networks have equal aspect ratios of the receptive fields, which leads to receptive fields either containing non-relevant information or being unable to fully cover the entire object. To this end, we propose Horizontal and Vertical Convolution, which is a plug-and-play module to address different aspect ratio problems. In our method, we introduce horizontal convolution and vertical convolution to expand the receptive fields in the horizontal and vertical directions, respectively, to reduce redundant receptive fields, so that remote sensing objects with different aspect ratios can achieve better receptive fields coverage, thereby achieving more accurate feature representation. In addition, we design an attention module to dynamically aggregate these two sub-modules to achieve more accurate feature coverage. Extensive experimental results on the DOTA and HRSC2016 datasets show that our HVConv achieves accuracy improvements in diverse detection architectures and obtains SOTA accuracy (mAP score of 77.60% with DOTA single-scale training and mAP score of 81.07% with DOTA multi-scale training). Various ablation studies were conducted as well, which is enough to verify the effectiveness of our model.