HCD-Mask: a multi-task model for small object detection and instance segmentation in high-resolution UAV images

Object detection and instance segmentation in images captured by unmanned aerial vehicles (UAVs) are becoming increasingly important. Although considerable progress has been made, there still exist challenges for the detection of small objects in high-resolution UAV images. Detecting objects and segmenting instances in aerial images are challengeable for extant methods due to the following two aspects: (1) original high-resolution images are usually resized before being fed into models. As a result, the features of small objects are mostly lost; (2) target objects like cars and trucks are very small in pixels, making them difficult to be distinguished from surrounding backgrounds. In order to address these issues, this paper presents a new multitask model for small object detection in high-resolution UAV images, namely HCD-Mask. Specifically, a high-resolution feature extractor is devised to distribute the features at different layers. Meanwhile, a coarse-to-fine region proposal network raises the intersection over union (IoU) threshold stage by stage, and maintains sufficient positive anchors in each stage. For more accurate mask estimation, the classification score and the position score are jointly scored for mask head. We demonstrate the effectiveness of our proposed model, by showing state-of-the-art performance on a UAV-based image dataset constructed by ourselves.