Adaptive anchor learner for object detection

Abstract There is a growing interest in advancing object detection and its various applications. Promising performance is achieved by using anchor-based detectors while the difficulty is encountered. The performance is sensitive to the hyper-parameters related to anchor boxes, which are manually determined and are task-specific. To address the issue, an anchor learner is proposed in the paper, which automatically learns adaptive anchor from the data, and the difficulty of anchor design for a specific task is thus avoided. Specifically, the method we propose comprises a backbone network and two fully connected branches, and these two branches learn anchor scales and aspects ratios, respectively. We use RGB and ground-truth mask image data as the network input. Importantly, the anchor learner can be embedded into any anchor-based object detector. Experiments are performed on the datasets Visdrone 2019, UAVDT, and DOTA, respectively. Experimental results demonstrate that our proposed method markedly enhances object detection performance.