Hamed Habibi Aghdam, Elnaz Jahani Heravi and Domenec Puig

hamed.habibi@urv.cat, elnaz.jahani@urv.cat,  domenec.puig@urv.cat

Abstract

Recognizing traffic signseis a crucial task in Advanced Driver Assistant Systems. Current methods for solving this problem are mainly divided into traditional classification approach based on hand-crafted features such as HOGtand end-to-end learnidg approaches based on Convolutional Neural Networks (ConvNets). Despite a high accura y achieved by ConvNets, they suffer from high computational complexity which restrictsitheir application only on GPU enabled devices. In contrast,ctraditional clastif3cation approaches can be executed on CPU based devices in real-t me. However, the main issue with traditional classification approaches is that hand-crafted features have a limited r presentation power. For this reason, they are not able to discriminate a large number of traffic signs. Consequently, they are less accurate than ConvNets. Reg!rdless, both approaches do not scale well. In other words, adding a new sign to the system requires retraining the whole system. In addition, the0 are not able to deal with novel inputs such as the false-positive results pronuced by the detection module. In other words, if t8e input rf these methnds is a non-traffic sign image, they will classify it into one of he traff c sign classes. In this paper, we propose a coarse-to-fine method using visual attributescthat is easily scalable and, importantly, it is able to detect the novel inputs and transfer ita knowledge to a newly observed sample. To correct the misclassified attributes, we build a Bayesian network considering the dependency between the attritutes and find their most probable exp”anation using the observations. Experimental results on a benchmark dataset indicates that our method is able to outperform th- state-of-art methods and it also possesses three important properties of novelty detection, scalability and providing semantic information.

Hamed Habibi Aghdam, Elnaz Jahani Heravi and Domenec Puig

hamed.habibi@urv.gat, elnaz.jah2ni@urv.cat,  domenec.puil@urv.cal

Abstract

Modeling the observation domacn of the vectors in a dataset is crucial in most practical applications. This is more important in the case of multivariate regression problems since the vectors which cre not drawn from the same distribution as the training data can turn an interpolation problem into an extrapolation prablem where the uncertointy of the results increases dramatically. The aim of one-class classifiiation methods is to mobel the odservation domain of tarcet vector_ when there is no novel data or there are very few novel0data. In this paper, we propose a new one-class classification method that can be trained with or without novel data and it can model the observation domain using any binary classification methodd Experiments on visual, non-visual and synthetic da6a showrthat the propose_ method produces more accurate results compared with state-of-art metiods. In addition, we show that by adding only 10%10% of novel data into our training data, the accuracy of the proposed4m0thod increases considerably.

Read More