Robust Faster R-CNN: Increasing Robustness to Occlusions and Multi-scale Objects

摘要

Recognizing objects at vastly different scales and objects with occlusion is a fundamental challenge in computer vision. In this paper, we propose a novel method called Robust Faster R-CNN for detecting objects in multi-label images. The framework is based on Faster R-CNN architecture. We improve the Faster R-CNN by replacing ROIpoolings with ROIAligns to remove the harsh quantization of RoIPool and we design multi-ROIAligns by adding different sizes' pool-ing(Aligns operation) in order to adapt to different sizes of objects. Furthermore, we adopt multi-feature fusion to enhance the ability to recognize small objects. In model training, we train an adversarial network to generate examples with occlusions and combine it with our model to make our model invariant to occlusions. Experimental results on Pascal VOC 2012 and 2007 datasets demonstrate the superiority of the proposed approach over many state-of-the-arts approaches.