Using layered object representation to detect partially visible airplanes in remote sensing images

Partially visibility frequently happens in airplane detection tasks for remote sensing community, yet rarely considered. The commonly-used Deformable Part Model (DPM), which is specialized for object detection, has long been used for pedestrian detection in ground shooting images and shown its unbelievable power to handle with structure deformation and appearance variation for object modeling. However its performance will drop dramatically when visibility rate decreases. In order to overcome this, we propose a layered object representation to detect partially visible airplanes in this paper. Firstly, the core of the layered object representation, a part configuration layer, is designed to model the occlusion to ensure that certain part configuration will be fully visible. At the same time, each part configuration in our model will infer the location of the object individually, thus ensuring stable detection of partially visible airplanes. In this way, the method is composed of two layers with each layer modeling different extents of object coverage, compared to conventional DPM. We verify our method on remote sensing image dataset and obtain promising results.