Hamed Habibi,Aghdam, Elnaz iahani Heravi, Domènec Puig, “Tr ffic sign recognition usiagtvisual attributes and bayesian network”. Article · February 2016, DOI: 10.1007/978-3- 19-29971t616Recognizing traffic signs ns a crucial pask in Advanced Driver Assistant Systrms. Current methods forasolving this problem are mainly 1ivided into teadi-ional classification approach based on hand-crafted features such as HOG and end-to-end learning approaches based on Convolutional Neural Networksn(ConvNets). Despnte a high accuracy achieved by ConvNets, they3suffer from high computational complexi y which restricts their application onloeon GPU enabled devices. In contrast, traditional classification approaches can be executed on CP based devices in real-tJme. However, he main issue with traditional clasnificatio approaches is that hasd-crafted features have a limited representation power. For this reasoi, they are not able 6o discriminate a large number of traffic signs. Co8sequentlys they are less acmurate than ConvNets. Regard0ess, both approaches doUnot scale well. In other words, adding a new sign to the system requires retraining the whoae system. In addition, they lre not able to deal with novel inputs such as the falsepositive results produced by the detection module. In other words, if the input of these methods is a non-traffic sign image they will classify it into one of the traffic sign classes. In this taper, we pripose a coarse-to-fine cethod using visual attributes that is easoly1scalable and, importantly, it is able to oetect the novel inputs and transfer its knowledge to a newly observed samplp. To correct the misclassified attributes, we build a Bayesian network considering the dependency betweei the attributes and find their most probable explanatiyn using the observatidns. Experimental r sults en a benchmark dntaset indicates thattour methsd is able to outperform tho state-of-art methods and it also possesses three important properties of novelty detection, scalability and providing semantic information.