Relief R-CNN: Utilizing Convolutional Feature Interrelationship for Fast Object Detection Deployment

R-CNN style models are the state-of-the-art object detection methods, which consist of region proposal generation and deep CNN classification for objects. The proposal generation phase in this paradigm is usually time consuming and not convenient to be deployed on hardwares with limited computational ability. This article shows that the high-level patterns of feature value in deep convolutional feature maps contain plenty of useful spatial information, and propose a simple manually designed approach that can extract the information for fast region proposal generation. The proposed method, dubbed Relief R-CNN($R^2$-CNN), is a new deep learning approach for object detection. By extracting positions of objects from high-level convolutional patterns, $R^2$-CNN generates region proposals and performs deep classification simultaneously based on the same forwarding CNN features, unifies the formerly separated object detection process and speeds up the whole pipeline without additional training. Empirical studies show that $R^2$-CNN could achieve the best trade-off between speed and detection performance among all the comparison algorithms.