Fast Multi Semantic Pyramids via Cross Fusing Inherent Features for Different-Scale Detection

Currently, there are many works exploring how to more fully and efficiently use multi-scale feature maps of deep convolutional neural networks to improve the performance of object detection. But most of these works are devoted to predicting, respectively, on multi-scale feature maps or blending multi-scale feature maps for enriching representation. In this paper, we present a new method of cross fusing feature, named multi-semantic pyramids (MSP), for detecting different-scale objects. Various scale objects are predicted, respectively, by corresponding semantic pyramids (SP), each SP can produce rich semantic features for predicting via reusing inherent multi-scale feature maps from the network backbone. Through promoting the reuse of inherent feature layers, our MSP can improve detection performance with marginal extra cost. In addition, since the reuse connection of the MSP facilitates the conduction of the gradient, the convergence of the network is greatly improved. The experimental results on the PASCAL VOC and COCO datasets illustrate that our MSP can achieve more competitive detection accuracy.