This letter studies the problem of air-to-air visual detection of micro unmanned aerial vehicles (UAVs) by monocular cameras. This problem is important for many applications such as vision-based swarming of UAVs, malicious UAV detection, and see-and-avoid systems for UAVs. Although deep learning methods have exhibited superior performance in many object detection tasks, their potential for UAV detection has not been well explored. As the first main contribution of this letter, we present a new dataset, named Det-Fly, which consists of more than 13 000 images of a flying target UAV acquired by another flying UAV. Compared to the existing datasets, the proposed one is more comprehensive in the sense that it covers a wide range of practical scenarios with different background scenes, viewing angles, relative distance, flying altitude, and lightning conditions. The second main contribution of this letter is to present an experimental evaluation of eight representative deep-learning algorithms based on the proposed dataset. To the best of our knowledge, this is the first comprehensive experimental evaluation of deep learning algorithms for the task of visual UAV detection so far. The evaluation results highlight some key challenges in the problem of air-to-air UAV detection and suggest potential ways to develop new algorithms in the future. The dataset is available at https://github.com/Jake-WU/Det-Fly.

Recent object detection networks suffer from performance degradation when training data and test data are distinct in image styles and content distributions. In this paper, we propose a domain adaptive method, Adversarial Mixing (AdvMix), where the label-rich source domain and unlabeled target domain are jointly trained by the adversarial feature alignment and a self-training strategy. To diminish the style gap, we design the Adversarial Gradient Reversal Layer (AdvGRL), containing a global-level domain discriminator to align the domain features by gradient reversal, and an adversarial weight mapping function to enhance the stability of domain-invariant features by hard example mining. To eliminate the content gap, we introduce a region mixing self-supervised training strategy where a region of the target image with the highest confidence is selected to merge with the source image, and the synthesis image is self-supervised by the consistency loss. To improve the reliability of self-training, we propose a strict confidence metric combining both object and bounding box uncertainty. Extensive experiments conducted on three benchmarks demonstrate that AdvMix achieves prominent performance in terms of detection accuracy, surpassing existing domain adaptive methods by nearly 5% mAP.