Multi-Label Classification of Traffic Scenes
Ivan Sikirić, Karla Brkić, Ivan Horvatin, Siniša Šegvić · Proceedings of the Croatian Computer Vision Workshop · 2014
This work deals with multi-label classification of traffic scene images.We introduce a novel labeling scheme for the traffic scene dataset FM2.Each image in the dataset is assigned up to five labels: settlement, road, tunnel, traffic and overpass.We propose representing the images with (i) bag-of-words and (ii) GIST descriptors.The bag-of-words model detects SIFT features in training images, clusters them to form visual words, and then represents each image as a histogram of visual words.On the other hand, the GIST descriptor represents an image by capturing perceptual features meaningful to a human observer, such as naturalness, openness, roughness, etc.We compare the two representations by measuring classification performance of Support Vector Machine and Random Forest classifiers.Labels are assigned by applying binary one-vs-all classifiers trained separately for each class.Categorization success is evaluated over multiple labels using a variety of parameters.We report good classification results for easier class labels (road, F 1 = 98% and tunnel, F 1 = 94%), and discuss weaker results (overpass, F 1 < 50%) that call for use of more advanced methods.