Indoor scene classification: A comparative study of feature detectors and local descriptors

Abstract

Indoor scene classification is a challenging problem in computer vision. In order to achieve an accurate solution for this task, a model that can exploit the discriminating information between different scene categories is necessary. In this paper, local feature extraction is suggested with supervised classification techniques for scene recognition in indoor environments. A comparative study between several feature detectors: SIFT, SURF, FAST, ORB, MSER, BRISK, and several local descriptors: SIFT, SURF, ORB, BRISK is presented. Two different classifiers, SVM and k- NN, are used for classification. The different techniques have been tested using the MIT dataset for indoor scenes and all the corresponding performance of each combination has been reported. Upon inspecting obtained results, it is observed that the combination of MSER detector with ORB as descriptor using SVM performs the best, as they give a combination of relatively high accuracy and low complexity; in terms of average execution time and memory space.