In most Brazilian capitals, there is a vast interest in safety and mobility improvements regarding the increasing number of vehicles circulating on the streets. These cities are investing more and more in security systems, including surveillance cameras which, over time, has far outnumbered the people available to monitor them. Furthermore, monitoring performed by humans has disadvantages such as inefficiency and high costs, when compared to autonomous systems. This is a significant problem because, besides these challenges, security footage is often analyzed reactively instead of proactively. This is to say that footage is first captured, opposite to being analyzed in real-time. As cities grow, the number of vehicles increase and, as consequence, also increase accidents, traffic jams, and traffic violations. Therefore, traffic management systems, particularly in medium and large cities, are facing increasingly complex challenges. Therefore, research on active traffic surveillance, i.e. monitoring and managing traffic flow, has attracted much attention recently.Obtaining vehicle information from images is a widely discussed topic in the literature. Within this context, for some applications, color recognition is one of the key challenges. A system that can efficiently and accurately recognize the color of a vehicle can be essential for systems regarding public and traffic safety. Possible applications are the localization of stolen vehicles, identification of vehicle registration fraud or illegal alterations to a vehicle. To classify a car image regarding its color, several difficulties have to be overcome. Although it may look like a simple task for the human eye, it is a fundamental problem in computer vision, since the apparent color of a vehicle changes as a function of time, space, and light conditions.
This dataset was created due to the difficulty to precisely recognise the color of a vehicle under variable light conditions. It can be used for testing image segmentation methods as well as training classifiers. The key point is, given an image with a vehicle, eliminate all elements other than the car body (background, glasses, wheels, licence plate, taillights, rubber/plastic trims and other elements, etc) that are not useful for idenfifying the color.
The dataset contains a total of 130 images, separated in training (100) and testing (30) sets. There are 10 classes, each one representing a basic color, according to the palette of Figure 1. For each basic color, three tone variants were added (light, regular, dark):
Related Papers: