Semantic recognition
Lotfi Abdi, Aref Meddeb · 2017
Understanding contents of an image, or scene labeling, is an important yet very challenging problem in artificial intelligence and computer vision to improve road safety. Semantic labeling and object detection in road scenes are strongly correlated tasks. Motivated by the complementary effect of the two tasks, we presented a novel framework to address the scene understanding problem. In this paper we propose a new framework for semantic labeling and object detection problem which is able to combine ideas from deep Convolutional Neural Network (CNN) for object detection and fully-connected Conditional Random Field (CRF) for segmenting and labeling. Specifically, we develop a new framework uses global image features to predict detection which drastically reduces its errors from background detections and a pairwise CRF is used as a post-processing step to enforce spatial consistency in the structured prediction. By combining the consistency between final detection results of CNN and CRF based graphical models, our unified framework can effectively leverage the advantages of leading techniques i.e., CRF and CNN for these two tasks. Extensive experiments on the PASCAL VOC 2007/2012 data sets demonstrate the effectiveness of our framework for Scene Understanding tasks.