Perceptual organization approach based on Dempster-Shafer theory

SUMMARY In this paper, we deal with the perceptual organization which is a crucial problem in computer vision. Perceptual organization consists in structuring the image into perceptually significant groups. These structures will be used in higher level processes such as image understanding or object recognition. The grouping is based on application of the Gestaltic perceptual phenomena. Originally, the Gestaltic theory tended to explain psychological and physiological mechanisms of the biological vision. In computer vision, perceptual organization provides complexity reduction for high level treatments and also tackles the problem of noise which induces fragmentation. However, its implementation is confronted to the combinatorial aspect of the process and the variability of the scene. Thus, we propose, in this paper, a perceptual grouping method based on the application of Gestaltic rules, by the Dempster-Shafer theory. The aim of this probabilistic approach is twofold. First, this method is applied in order to rectify segmentation mistakes by restoring the coherence of the primitives. Next, it allows to form groups of primitives which correspond to separate real objects in the scene. The method, which is proposed in this paper, is completely bottom-up and is able to discern different objects in the scene without prior knowledge. Moreover, it uses no threshold and thus, the variability of the scene does not affect the final results. We show in this paper how we apply the Dempster-Shafer theory, usually used in data fusion, in order to obtain an optimal adequation between our problem and this tool. In order to demonstrate the robustness and the reliability of our algorithm, we finally show experimental results on both real indoor and outdoor scenes. ABSTRACT In this paper, we propose an application of the perceptual organization based on the Dempster-Shafer theory. This method is divided into two parts which respectively rectifies the segmentation mistakes by restoring the coherence of the segments and detects objects in the scene by forming groups of primitives. We show how we apply the Dempster-Shafer theory, usually used in data fusion, in order to obtain an optimal adequation between the perceptual organization problem and this tool. We show that without any prior knowledge and any threshold, our bottom-up algorithm detects efficiently the different objects even in cluttered environment. Moreover, we demonstrate its robustness and flexibility on indoor and outdoor scenes without any modification of parameters.