StairsNet: Mixed Multi-scale Network for Object Detection

It is common to choose image classification network as backbone in the object detector. The art-of-the-state image classification network exhibits excellent performance on image classification, but that network hurts the detection efficiency, mainly due to the coarseness of features from several convolution and pooling layers. In this paper, we present a single deep neural network with inceptions, called StairsNet, to take advantage of the art-of-the-state image classification network in object detection. In contrast to previous single network SSD [13] which uses VGG-16 as a feature to extract network, our approach applies recently state-of-the-art classification network Residual Network (ResNets [5]). Meanwhile, to avoid coarseness of the last CNN feature, StairsNet not only utilizes various of scale features, but also mixes different scale features to predict. To this end, we insert two stairs-like architectures into the network: top stairway network that mixes multiscale feature maps as input to predict bounding boxes and bottom stairway network that turns into two different scale feature branches. Our StairsNet significantly increases the PASCAL-style mean Average Precision (mAP) from 75.0% (SSD + ResNet-101) to 77.7%. Code is available at https://github.com/gwyve/caffe/tree/StairsNet.