Sequential Feature Fusion for Object Detection

In an image, the category and the location of an object are related to global, spatial and contextual visual information of the object, which are all extremely important for accurate and efficient object detection. In this paper, we propose a region-based detector named Sequential Feature Fusion Network (SFFN) which simultaneously utilizes global, spatial and multi-scale contextual Region-of-Interest (RoI) features of an object and fuses them by a novel method. Specifically, we design a Feature Fusion Block (FFB) to fuse global and multi-scale contextual RoI features, which are extracted by RoI pooling layer. Then we apply the concatenation operation to integrate the fused feature with spatial RoI feature extracted by Positive-Sensitive RoI (PSRoI) pooling layer. The experimental results show that the performance of SFFN obtains significant improvements on both the PASCAL VOC 2007 and VOC 2012 datasets.