Comparison between Bayesian network classifiers and SVMs for semantic localization. (1st December 2016)

• Record Type: Journal Article
• Authors: Rubio, Fernando; Martínez-Gómez, Jesus; Julia Flores, M.; Puerta, José M.
• Is Part Of: Expert systems with applications. Volume 64(2016)
• Page Start: 434
• Page End: 443
• Publication Date: 2016-12-01
• Subjects: Semantic localization -- Bayesian networks -- Robotics -- Robot vision -- Structural learning -- Indoor scene classification; Expert systems (Computer science) -- Periodicals; Systèmes experts (Informatique) -- Périodiques; Electronic journals; 006.33
• DOI: 10.1016/j.eswa.2016.08.029
• Languages: English
• ISSNs: 0957-4174

Abstract

Highlights: We propose a method for the use of image descriptors as input in graphical model. We compare the results obtained in two datasets when using SVMs and BNCs. A large number of inputs decreases the results of BNCs with structural learning. Graphical models provide capabilities of an easier interpretation than SVMs. BNCs allow incremental modelling, so new information can be added to enrich them. Abstract: This work presents a methodology to apply Bayesian networks classifiers (BNCs) to the problem of semantic localization in robotics. This task consists of determining where the robot is located by using semantic annotations instead of metric locations, and based on robots perceptions, namely images. The proposal covers the two key steps of (1) extracting descriptive features from the input image and (2) construction and evaluation of models, comparing the performance of BNCs technologies with SVMs solutions. The experimentation is performed over two different datasets, and the results, given in terms of accuracy, provide a quite appealing analysis where specialization versus generalization or model complexity are considered. Overall BNCs proved to be quite competitive, and appear to be a very promising tool for future applications since they would allow the introduction of additional contextual information to the processing pipeline.