APLNet: Attention-enhanced progressive learning network

Hui Zhang, Danqing Kang, Haibo He, Fei-Yue Wang

Abstract: Highlights•We propose APLNet using multiple stages for progressive detection to improve the performance of single-stage detectors. In each stage, only a convolutional layer is applied to increase the generalization capability and keep high efficiency in the regression network.•We design an attention enhancement module to generate the attention map for injecting semantic information to the features of the low-level layer. The attention map is achieved under the guidance of boxes-induced segmentation annotations. By embedding this attention map into the low-level features, it can obtain more semantically meaningful information without additional pixel-wise segmentation annotations.•Experiments on PASCAL VOC 2007, PASCAL VOC 2012 and MS COCO datasets demonstrate that the proposed APLNet achieves competitive performance both in accuracy and efficiency.