A one-stage object detector based on SSD, named FIFENet (Feature Integration and Feature Enhancement Network), is proposed in this paper to settle the deficiency of SSD in small objects detection. Two blocks are designed in FIFENet: a feature integration block and a feature enhancement block. Feature integration block fuses the feature map in shallow layers to improve the performance on small objects. Feature enhancement block adopts the residual network (Res2Net) and attention mechanism to enhance feature integration. Experimental result shows that the mean average precision (mAP) on PASCAL VOC2007 data set is 3.1% higher than vanilla SSD, and the accuracy improvement on birds, bottles, chairs, and plants are 3.6%, 9.5%, 5.4%, and 5.5% separately. Results demonstrate that the FIFENet can achieve high detection accuracy while maintaining real-time performance.
|