Revisiting Faster R-CNN: A Deeper Look at Region Proposal Network

Currently, state-of-the-art object detectors are based on Faster R-CNN. We firstly revisit Faster R-CNN and explore problems in it, e.g., coarseness of feature maps for accurate localization, fixed-window feature extraction in RPN and insensitivity for small scale objects. Then a novel object detection network is proposed to address these problems. Specifically, we utilize a two-stage cascade multi-scale proposal generation network to get high accurate proposals: an original RPN is adopted to initially generate coarse proposals, then another network with multi-layer features and RoI pooling layer are introduced to refine these proposals. We also generate small scale proposals in the second stage simultaneously. After that, a detection network with multi-layer features further classifies and refines proposals. A novel 3-step joint training algorithm is introduced to optimize our model. Experiments on PASCAL VOC 2007 and 2012 demonstrate the effectiveness of our network.