Empirical Study of RepPoints Representation for Object Detection in Aerial Images

Anchor-based detectors have dominated object detection for several years. These rely heavily on rectangular bounding boxes representation, which is convenient to use but reveals severe limitations, causing the inaccurate location of the objects with dense distribution or arbitrary direction. In this paper, we first utilize a new finer representation of objects, RepPoints (representative points), for improved feature extraction and object localization on aerial traffic images. Then, we experimentally fine-tuned the RPDet – an anchor-free object detector based on RepPoints – to prove that this approach can achieve the same effective performance as the state-of-the-art anchor-based detection methods. Our experimental modifications include the adoption of advanced models such as ResNet50, ResNeXt101 and Res2Net101 as backbone; Besides, we implement modules of DCN (Deformable Convolution Networks) for backbone architecture. To the best of our knowledge, the modified system is the current best performer on the task of object detection with 23.6% in AP and 42.8% in AP50 on the VISDRONE-DET detection benchmark.