Perceptual Grouping Based on Iterative Multi-scale Tensor Voting

Abstract

Abstract. We propose a new approach for perceptual grouping of oriented segments in highly cluttered images based on tensor voting. Segments are represented as second-order tensors and communicate with each other through a voting scheme that incorporates the Gestalt principles of visual perception. An iterative scheme has been devised which removes noise segments in a conservative way using multi-scale analysis and re-voting. We have tested our approach on data sets composed of real objects in real backgrounds. Our experimental results indicate that our method can segment successfully objects in images with up to twenty times more noise segments than object ones.

An Automatic Framework for Figure-Ground Segmentation in Cluttered Backgrounds

“…The only free parameter is the scale factor σ , which defines the voting fields. Specific details can be found in [2] and [4] Although the scale factor σ of the tensor voting framework has been shown to be fairly stable [6], it is subjected to the same trade-off of every scale-dependent method: small scales capture local structures while large ones capture global configurations. In real scenarios, however, structures emerge from different scales and the prediction of the optimal scale is fairly complicated.…”

“…Moreover, it is impossible in general to choose a fixed threshold that would provide a good figure segmentation due to the complexity and amount of background in images. We have addressed both issues in [4] by introducing a multi-scale scheme that removes background segments conservatively in an iterative fashion, leading to an improved figure-ground segmentation.…”

“…Like in [4], we eliminate image edges iteratively by analyzing their behavior at multiple scales, however, the main advantage of our method is that it works in an automatic fashion. To halt the iterative process automatically but also to determine at each iteration what action is more effective for a certain type of histogram, we have devised a methodology that categorizes saliency histograms in different cases.…”

“…Tensor voting represents input data as tensors and interrelates them through voting fields built from a saliency function that incorporates the Gestalt principles of proximity and good continuation. In the past, we have employed tensor voting in order to characterize the salience of edge segments [4]. In particular, we performed figure-ground segmentation by analyzing saliency information at multiple scales and removing low-saliency segments in an iterative fashion.…”

“…Following Loss et al [4], we perform figure-ground segmentation by analyzing the saliency of edge segments at multiple scales and removing low-salience segments using an iterative scheme. The main contribution of our work is automating the above framework by introducing a case-based thresholding scheme to eliminate low-saliency segments more effectively and criteria for stopping the iterative voting without user intervention.…”