Local Adjustment Block Net for Object Detection

Abstract

Recently, object detectors based on Receptive Fields (RFs) in human visual systems have stronger feature extraction capabilities and have achieved great detection performance, such as Inception, ASPP and RFBNet. However, while having capabilities to extract more contextual information, these detectors also capture redundant information, which will reduce the precision of detection. In this paper, we propose a novel and lightweight block based on spatial attention mechanism to solve this problem effectively. Compared with RFB, it can better capture effective contextual information in the feature map and suppress redundant information. Moreover, we propose a local enhancement strategy, which can sparsely locate regions that contains rich feature information and enhance them locally. Experimental results show that our proposed method gains 0.6% mAP improvement on the PSACAL VOC dataset and 0.5% mAP improvement on the COC02019 test-dev set.