Knowledge distillation(KD) is a widely-used technique to train compact models in object detection. However, there is still a lack of study on how to distill between heterogeneous detectors. In this paper, we empirically find that better FPN features from a heterogeneous teacher detector can help the student although their detection heads and label assignments are different. However, directly aligning the feature maps to distill detectors suffers from two problems. First, the difference in feature magnitude between the teacher and the student could enforce overly strict constraints on the student. Second, the FPN stages and channels with large feature magnitude from the teacher model could dominate the gradient of distillation loss, which will overwhelm the effects of other features in KD and introduce much noise. To address the above issues, we propose to imitate features with Pearson Correlation Coefficient to focus on the relational information from the teacher and relax constraints on the magnitude of the features. Our method consistently outperforms the existing detection KD methods and works for both homogeneous and heterogeneous student-teacher pairs. Furthermore, it converges faster. With a powerful MaskRCNN-Swin detector as the teacher, ResNet-50 based RetinaNet and FCOS achieve 41.5% and 43.9% $mAP$ on COCO2017, which are 4.1% and 4.8% higher than the baseline, respectively.

Deep learning-based methods are continuously boosting the performance of detecting objects in natural images. On the contrary, detecting objects in Unmanned Aerial Vehicle (UAV) images remains to be a difficult task in the field of computer vision, due to the challenge of training a well-performed detection model working on UAV images which usually contain instances with varied orientations, scales, and contours, etc. Furthermore, only a few researchers have focused on this field, probably because of difficulties in UAV data acquisition and labelling. Inspired by this, we collected a large-scale dataset with multi-scale and high-resolution UAV images, named MOHR, which contains 10,631 images captured by a UAV affixed with three kinds of cameras. Since these images were captured in a suburban environment, we manually annotated five classes of objects, including car, truck, building, collapse and flood damage. An empirical study was then conducted by adopting six advanced object detection methods all of which are based on deep learning technologies. The results indicate the great potential of these evaluated object detection models, but also reveal that the research on such a challenging UAV dataset using current deep learning techniques is far reaching.