Participants: 'Ajay Mishra', Ching Teo, Cornelia Fermuller, Yiannis Aloimonos

To segment foreqround objects from the background, we used the algorithm developed in (Mishra et al 09). This algorithm, in the absence of predefined color or texture models formulates segmentation by finding the closed boundary curve separating foreground from background. In a nutshell the algorithm works as explained visually in Fig 1. The vision system fixates at a point (or selects a point in the image) and computes an edge map. To achieve size-invariance, the boundary edge map is transformed from the Cartesian coordinate system to the polar coordinate system with the fixation point as the pole. In the polar coordinate system, we then find the contour enclosing the object. This can be formulated as an energy minimization in a conditional random field (CRF), and the minimization is solved by the graph cut algorithm. In summary, the algorithm selects a point in the image (the fixation point) and finds a closed contour containing that point. Edges are computed by combining the RGB images with the depth maps as follows: First edges are located in the intensity images using the probabilistic edge detector in (Martin et al.04). Then we check for every edge, whether in the depth map there is a corresponding discontinuity and strengthen the probabilistic value of the edges, which have nearby depth discontinuities. The process is illustrated in Figure 1.

Figure 1: Illustration of the segmentation algorithm

Figure 2 illustrates the segmentation results.

Figure 2: Segmentation results


  1. Martin, C. Fowlkes, and J. Malik (2004) Learning to detect natural image boundaries using local brightness, color and texture cues, T-PAMI, 26(5):530–549, May 2004.
  2. Mishra A, Y. Aloimonos, and LF. Cheong (2009), Active Segmentation with Fixation, ICCV, 2009.
  3. Mishra, C. Fermüller , and Y. Aloimonos . Active Segmentation for Robotics, IROS, 2009.