Experimental results

Figure 3: Background subtraction algorithm. (a) and (b) left and right images of the background. (c) Left image with object. Bottom row: results using (d) simple tolerance, (e) simple tolerance on refined disparity map, and (f) windowed means on refined disparity map.

The results of our algorithm are shown in Figure 3. The first pair of images, Figures 3a and 3b show two camera views of an empty scene; Figure 3c is the key image at run time with the person in the view of the camera. The background is subtracted by one of the following methods (Figure 3d-f): 1) simple tolerance range ${\hbox{\rlap{$\epsilon$}\kern .4pt{$\epsilon$}}}$, 2) simple tolerance range ${\hbox{\rlap{$\epsilon$}\kern .4pt{$\epsilon$}}}$ on a disparity map, adjusted by the refinement procedure, and 3) windowed means on a refined disparity map. Note the presense of the occlusion shadow right behind the person.

Figure 4: Removing the occlusion shadow. a) left reference camera view, b) key camera view, c) right reference camera view, d) subtraction using a) and b) views, e) subtraction using b) and c) views, f) removing of the shadows using both reference cameras.

Figure 4 demonstrates the process of removing the occlusion shadow. Three camera views are shown in the top row, where Figure 4a is a left reference camera view, Figure 4b is a key camera view and Figure 4c is a right reference camera view. Figures 4d and 4e show subtraction performed on two pairs of images. Figure 4f shows the result of using all three cameras.