Scalable scene understanding using saliency-guided object localization

Given an image, scene understanding is the process of segmenting and identifying the objects present, and classifying the overall scene. Several frameworks already exist to perform these tasks coherently but training of their probabilistic models is time consuming thereby limiting their scalability. This paper presents a scalable framework adopting an object-based approach. The steps taken by the algorithm are saliency detection for unsupervised object discovery, graph-cut for object segmentation, bag-of-features for object classification and binary decision trees for scene classification. A region of interest (ROI) detector is proposed to automatically provide object location priors from saliency maps for graph-cut. We tested our system on a novel NUS/NTU dataset and compared the scene classification accuracy using different classifiers. Unlike other existing frameworks, the proposed algorithm is scalable and can easily accommodate more object and scene classes.