Oriented Object Detection for Large Aspect Ratio Vehicles in Remote Sensing Images

Abstract

Existing vehicle detection methods in remote sensing images encounter challenges when detecting vehicles with large aspect ratios. Due to the big scale gap between the long edge and the short edge, large aspect ratio vehicles are hard to extract fine features. In addition, large aspect ratio results in strong orientation information and the inconsistency between regression task and classification task is even more severe. To address these issues, this paper proposes a Large Aspect Ratio Vehicles Detector (LARDet). Aiming at the difficulty of feature extraction for objects with large aspect ratios, we adopt more data augmentation and introduce PAN structure to pass through the short edge feature from shallow layer to deep layer, so as to extract more discriminative features. A lightweight Boxes Quality Predication Module (BQPM) is designed to alleviate the inconsistency between classification score and location accuracy. To alleviate the feature inconsistency between regression and classification, we further design the Align Classification Module (ACM), change the regression branch and classification branch from parallel to serial, then apply AlignConv to extract rotation-invariance feature for classification. A Large Aspect Ratio Vehicles Dataset (LAR1024) is proposed to evaluate our method. Compared with other SOTA methods, LARDet gains 5.0% AP on LAR1024 with the fastest speed of 23.9 FPS, which achieves a better speed-accuracy trade-off in the detection of large aspect ratio vehicles.