Abstract
Object detection attracts much attention for its great value in theories and applications. The one-stage single shot object detectors outperform the two-stage methods in running speed with a comparable performance. In this paper, we propose three novel strategies, to further improve the performances of single shot detector without sacrificing their runtime efficiency. Firstly, we design the multi-scale context aggregation module to embeds the context information into the learned features. Secondly, we design the multi-path predicting head, which decouples the network layers and can easily learn the effective receptive fields of different aspect ratios, to detect objects of various aspect ratios better. Thirdly, we adopt a top-down feature map pyramid to detect objects using features of different semantic powers and resolutions. Sufficient ablation experiments are conducted to prove the efficiency of the proposed methods. We design a one stage single detector named as ISSD, using the three strategies. Experimental results on PASCAL VOC 2007 and 2012 shows ISSD achieves the new state-of-the-art on accuracy with the comparable running speed.