Joint Anchor-Feature Refinement for Real-Time Accurate Object Detection in Images and Videos

Object detection has been vigorously investigated for years but fast accurate real-world scenes remains a very challenging problem. Overcoming drawbacks of single-stage detectors, we take aim at precisely detecting objects static and temporal in real time. Firstly, as dual refinement mechanism, novel anchor-offset is designed, which includes an anchor refinement, feature location deformable head. This new mode able to simultaneously perform two-step regression capture object features. Based on the detection, network (DRNet) developed high-performance where multi-deformable head further designed leverage contextual information describing objects. As videos, networks (TRNet) (TDRNet) are by propagating across We also propose soft strategy temporally match motion with previous refinement. Our proposed methods evaluated PASCAL VOC, COCO, ImageNet VID datasets. Extensive comparisons verify superiority DRNet, TRNet, TDRNet. Consequently, our approaches run fairly speed, meantime achieve significantly enhanced accuracy, i.e., 84.4% mAP VOC 2007, 83.6% 2012, 69.4% 2017, 42.4% AP COCO. Ultimately, producing encouraging results, applied online underwater grasping autonomous system. Codes publicly available https://github.com/SeanChenxy/TDRN.