Dynamic Up-Sampling Network for Small Airplane Landing Gear Detection

Airplane landing gear detection is essential in the intelligent flight safety assurance system. However, it is challenging due to the small object scale. Modern detectors based on deep convolutional neural networks (CNNs) fail to detect landing gears accurately in real-time. In this work, an analysis of small object detection is presented. We demonstrate that the low discrimination of features and low density of anchors on the deep feature map make the CNN-based detectors failing to detect small objects accurately. Therefore, we propose increasing the feature discrimination and the anchor density for higher detection accuracy by increasing image resolution and exploiting the inherent multi-scale features of CNNs. To optimize the high computational overhead brought by increasing image resolution, we propose a simple and effective Dynamic Up-sampling Network (DUN) which dynamically up-samples the interesting regions that may have landing gears and uses a light weight one-stage detector (LWOD) for detection. To evaluate the effectiveness of our method, we compare DUN with the state-of-the-art CNN-based detectors. The results show that DUN achieves a competitive accuracy and an improved speed on detecting small landing gears. Specifically, DUN can run in 24 ms at 97.2 AP on a 2080Ti GPU. Therefore, it is a practical and accurate solution to small landing gear detection.