Decoupled Feature Pyramid Learning for Multi-Scale Object Detection in Low-Altitude Remote Sensing Images
Haokai Sun, Yaxiong Chen, Xiongbo Lu, Shengwu Xiong
Abstract
Recently, low-altitude remote sensing platforms are widely used for various practical applications. Object detection is a basic and significant technology, serving them. The scale imbalance problem is predominant in low-altitude remote sensing images which brings a great challenge to detect objects from these imageries. Consequently, in this paper, we boost performance from the perspective of mitigating scale imbalance issue. Firstly, we choose a one-stage object detector with decoupled heads as baseline because of its comparatively high efficiency and accuracy. Current decoupled heads ignore inter-layer relationship and the information contained. On the other hand, all existing feature pyramid structures generate one feature map for two branches at every layer. Inspired by them, we propose a novel feature pyramid network paradigm—Decoupled Feature Pyramid Network (DFPN) with consideration of different preferences for classification and localization. Meanwhile, introduction of feature pyramid architecture will cause performance deterioration of larger objects because upper layers receive insufficient supervision in the training phase. Therefore, we adopt a distinct supervision strategy—Level Supervision, which pays more attention to upper layers. We demonstrate extensive experiments on two popular benchmarks of object detection in low-altitude remote sensing images to validate the effectiveness of our proposed method. Whats more, we introduce a Scale Imbalance metric to quantify the degree of size change of objects to better illustrate the ability of relieving scale imbalance problem. Finally, our proposed approach achieves state-of-the-arts performance on both datasets.