Research On Localization Accuracy In Visual Object Detection And Tracking

In computer vision,visual object detection and tracking are the fundamental and important research directions,which have also been widely used in industry.Visual object detection and tracking can provide the basic perception of surrounding objects,which can help other high level understanding tasks.In visual object detection and tracking,localization accuracy of objects is the key of the performances of detection and tracking results.In visual object detection,from a perspective of localization-quality,we find a vital contradiction that classification confidence can not fully reflect localization quality.To ease the contradiction,we propose a novel localization-quality estimation embedded detector,called LED.LED builds localization-quality estimation branch based on a single shot object detector,to estimate the localization confidence of each detection bounding box.In the inference phase,we propose to fuse the localization confidence with the classification confidence for further non-maximum suppression process.Comprehensive experiments are conducted on PASCAL VOC and KITTI datasets to demonstrated the effectiveness of LED with a little drop on running speed.Furthermore,in visual object tracking area,we find that current visual object trackers do not provide specific optimization in some challenging tracking scenes.To improve the localization accuracy in these challenging scenes such as scale change,shape deformation and occlusion,we propose the Corner Aided Tracker with Deep Regression Network,called CAT.In CAT,to obtain more accurate tracking bounding boxes,we regard the bounding box localization problem as a key point estimation problem.Through the localization,adaptive selection and fusion of bounding box corner key points as well as center key point,CAT can obtain accurate and robust tracking result bounding boxes,especially in challenging scenes such as scale change,shape deformation and occlusion.In comprehensive experiments on OTB,UAV123,La SOT and VOT datasets,CAT can provide state-of-the-art performances with high localization accuracy.