VistrongerDet: Stronger Visual Information for Object Detection in VisDrone Images
Junfeng Wan, Binyu Zhang, Yanyun Zhao, Yunhao Du, Zhihang Tong
Abstract
Existing methods are especially difficult to detect objects accurately in videos and images captured by UAV. In the work, we carefully analyze the characteristics of VisDrone DET 2021 dataset, and the main reasons for the low detection performance are tiny objects, wide scale span, long-tail distribution, confusion of similar classes. To mitigate the adverse influences caused thereby, we propose a novel detector named VistrongerDet, which possesses Stronger Visual Information. Our framework integrates the novel components of FPN level, ROI level and head level enhancements. Benefitted from the overall enhancements, VistrongerDet significantly improves the detection performance. Without bells and whistles, VistrongerDet is plug-gable which can be used in any FPN-based two-stage detectors. It achieves 1.23 points and 1.15 points higher Average Precision (AP) than Faster R-CNN and Cascade R-CNN on VisDrone-DET test-dev set.