ECascade-RCNN: Enhanced Cascade RCNN for Multi-scale Object Detection in UAV Images
Qizhang Lin, Yan Ding, Hong Xu, Wenxiang Lin, Jiaxin Li, Xiaoxiao Xie
Abstract
Due to the change of flight altitude and attitude of UAV, the object scale in UAV images exists difference which leads to a great challenge for object detection and has drawn wide attention. In this paper, an improved object detection network named ECascade-RCNN is proposed to deal with the multi-scale problem in object detection task for UAV images. We present an innovative Trident-FPN backbone to extract features and design a new attention mechanism to enhance the performance of the detector. Moreover, k-means algorithm is adapted to generate anchors so that the detection model can get better regression accuracy. We evaluate the proposed ECascade-R-CNN on Visdrone dataset through several ablation experiments and the results show that the ECascade-RCNN given in the paper is effective. The ECascade-RCNN is also used in the Visdrone2020 challenge and ranked 8th on the object detection track.