VisDrone-DET2020: The Vision Meets Drone Object Detection in Image Challenge Results

Abstract

Recently, automatic visual data understanding from drone platforms becomes highly demanding. To facilitate the study, the Vision Meets Drone Object Detection in Image Challenge is held the second time in conjunction with the 17-th International Conference on Computer Vision (ICCV 2019), focuses on image object detection on drones. Results of 33 object detection algorithms are presented. For each participating detector, a short description is provided in the appendix. Our goal is to advance the state-of-the-art detection algorithms and provide a comprehensive evaluation platform for them. The evaluation protocol of the VisDrone-DET2019 Challenge and the comparison results of all the submitted detectors on the released dataset are publicly available at the website: http: //www.aiskyeye.com/. The results demonstrate that there still remains a large room for improvement for object detection algorithms on drones.

The State of Aerial Surveillance: A Survey

“…5. The Cascade R-CNN [17] drops the performance by 50% in the aerial VisDrone [187] dataset compared with the ground-based COCO and Pascal VOC datasets [36]. The Faster R-CNN [117] also drops the performance by 30% in the aerial TinyPersons [173] dataset compared with the ground-based COCO and Pascal VOC datasets [171].…”

“…Multiple challenges have been organized in top-tier conferences. The challenges "Vision meets drones" [36] including detection, tracking and crowd counting have been organized in the last four consecutive years in ECCV 2018, ICCV 2019, ECCV 2020 and ICCV 2021 using the VisDrone dataset. The TinyPerson challenges [171], organized in ICCV 2019 and ECCV 2020, focus on the persons from a very long distance with a wide view using the TinyPerson dataset.…”

“…Many recent approaches [2], [148], [175] achieved good performance with Cascade R-CNN [17]. In fact, many entries, including the winning and runner-up entries, of the VisDrone detection challenges combined Cascade R-CNN [17] with other networks to achieve high detection performance [36]. Similarly, two of top-three detectors in the TinyPersons challenge employed Cascade R-CNN [171].…”

“…The scale of object resolutions also varies largely from 10 1 to 10 3 pixels not just within the dataset, but also within a single image [20]. There are three key approaches to deal with small object detection in aerial data: (i) improve feature maps for small objects [86], [138], (ii) incorporate context information of small objects [82], [97], and (iii) data augmentation [36].…”

“…However, this requires a very large scale dataset to cover all possible variations. Data augmentation [36], [171] and generative models such as GANs [22] can be used to generate extra data to deal with this challenge during training.…”