Existence Detection of Objects in Images for Robot Vision Using Saliency Histogram Features Christian Scharfenberger 1 , Steven L. Waslander 2 , John S. Zelek 1 and David A. Clausi 1 1 Systems Design Engineering, University of Waterloo, Canada 2 Mechanical and Mechatronics Engineering, University of Waterloo, Canada {cscharfenberger, stevenw, jzelek, dclausi}@uwaterloo.ca Abstract—In robotics and computer vision, saliency maps are frequently used to identify regions that contain potential objects of interest and to restrict object detection to those regions only. However, common saliency approaches do not provide information as to whether there really is an interesting object triggering saliency and therefore tend to highlight needless background as potential regions of interest. This paper ad- dresses the problem by exploiting histogram features extracted from saliency maps to predict the existence of interesting objects in images and to quickly prune uninteresting images. To validate our approach, we constructed a database that consists of 1000 background and object images captured in the working environment of our robot. Experimental results demonstrate that our approach achieves good detection performance and outperforms an existing existence detection approach [1]. Keywords-Existence detection, histogram features, saliency maps, robotics I. INTRODUCTION Image processing is increasingly gaining importance for applications in robotics. Examples include visual perception and servoing, object detection and recognition, and obstacle avoidance among others. Images captured in the working environment of a mobile robot contain a huge amount of visual information that can be subdivided into task-relevant and task-irrelevant. Extracting task-relevant information is of great importance in robotics, and involves the enhancement of image regions that contain task relevant information (e.g., interesting objects), and the inhibition of irrelevant regions such as background. As an example, the performance of object detection can be improved significantly when applied to regions containing interesting objects only. Hence, saliency detection for information selection in images has attracted a lot of attention. Saliency detection is the concept of highlighting those regions in images which stand out with respect to their neighborhood and which are unique in attributes – such as color, texture, etc. – relative to other regions. This forms an important first step in various tasks such as image segmentation [2], object detection [3], and object recognition [4]. Although the generation of saliency maps is an intensively studied field in image processing and robotics, most saliency approaches do not check whether there really exist inter- esting objects in images that trigger saliency measures. For example, robots performing search tasks in large areas rarely see objects of interest. In the case of images containing back- ground only, existing saliency approaches tend to highlight regions with unique attributes, yet irrelevant information. These regions may lead to large false positive detection rates when fed into object detection approaches. As such, we are interested in quickly pruning uninteresting images contain- ing only background information, and in keeping those that are crucial to the task of a robot. The challenging aspect associated with explicitly pruning uninteresting images is the choice of an appropriate representation for distinguishing whether there exist interesting objects in a saliency map or not. Prior work has attempted to address this issue by segment- ing saliency maps using thresholds and by keeping pixels whose saliency is above a given threshold (e.g., [5],[6]). The approaches presume that pixels associated with interesting objects have high, and pixels associated with background low saliency values. However, finding appropriate thresh- olds for suppressing background images is difficult since pixels associated with cluttered background may have strong saliency values as well. As an alternative, the approach of Wang et al. [1] predicts the existence of interesting objects in thumbnail images by exploiting global features and geometrical information from multiple saliency maps. This is based on the assumption that most thumbnail images contain single objects located in the center of an image. However, this assumption might not be valid for applications in robotics where multiple objects may be spread over the entire image, or a single object may be located at one image border. Therefore, an efficient method to quickly judge the existence of interesting objects based explicitly on features that do not rely on geometrical information or on a priori information would be of great interest. The main contribution of this paper is the introduction of a simple yet effective approach to existence detection based on global features extracted from the probability distribution function (PDF) of saliency maps. A histogram of saliency values (PDF) is constructed to effectively remove geometrical information such as the object location in the saliency map data. By exhibiting the largest variance across 2013 International Conference on Computer and Robot Vision 978-0-7695-4983-5/13 $26.00 © 2013 IEEE DOI 10.1109/CRV.2013.25 75