Multi-scale Pyramid Feature Maps for Object Detection

this paper presents how we can achieve the state-of-the-art accuracy in multi-scale objection detection task while adopting and combining the recent technical innovation in deep learning. Following the common pipeline of CNN feature extraction, we mainly design the architecture of feature extraction which exploits the idea of feature pyramid. We further add an extra 1*1 convolution layer to benefit feature extraction, via the batch normalization. In addition, the designed network architecture for feature extraction combines low-resolution and high-resolution feature layers to predict the category of the object in images. The new architecture is trained with the help of batch normalization, mean pooling based on plateau detection. The proposed architecture shows competitive results compared to some state-of-the-art algorithms both in accuracy and in speed on some datasets.