Dimensionality reduction for off-line object recognition and detection using supervised learning

Abstract

Object recognition and detection is an area of study, within intelligence and computer vision. It finds applications in fields such as surveillance, detailed activity analysis, robotics and object tracking. The primary focus of research papers in this domain revolves around enhancing the precision of object identification and detection regardless of whether the objects are located indoors or outdoors. To address this challenge, a new approach involving the utilization of SIFT features for information extraction has been proposed. Our approach encompasses two components; the implementation of dimensionality reduction through principal component analysis (PCA) to eliminate redundancies; secondly the incorporation of feature vector encoding using fisher encoding techniques. The RGB-D dataset employed contains 300 objects across scenarios with emphasis on colored aspects rather than depth. The SIFT features are categorized using a support vector machine (SVM) into 7 classes. When compared to using SIFT features integrating them with encoding methods notably enhances recall, precision and F1-score by than 30% through fisher encoding and PCA techniques. The study concludes with an evaluation based on n-cross validation methodology along, with detailed experimental results.