Robust real-time aircraft detection with multi-task cascaded calibration networks

Abstract

Aircraft detection is notoriously challenging owing to the orientation and size variations of aircraft objects. Existing detection pipelines compromise with efficiency or accuracy to deal with the large visual variations. We present a novel cascaded framework that joins object detection and orientation prediction through multi-task learning. The cascaded framework consists of three stages and operates in a coarse-to-fine manner. Each stage simultaneously rejects false targets, regresses the locations of object candidates, and calibrates the orientations of the candidates to upright gradually. After each stage, the range of orientation angles continuously decreases, which contributes to accurately distinguishing aircraft from non-aircraft. In addition, we design a generative-adversarial architecture in the cascaded framework. Specifically, its generator learns to cheat the discriminator by producing the feature representation that is invariant to the change of aircraft size. Meanwhile, the discriminator competes with the generator and aims to accurately discriminate the feature representations of small objects from the representations of large ones. Through competition between the generator and the discriminator, the feature distributions over small and large objects are made similar, thus resulting in more accurate detection. Experimental results demonstrate our method achieves significant improvements over existing algorithms on the challenging detection dataset, while still keeps real-time performance.