Stacked Pebbles Stacked Cubes 3 Experimental Results Human Face 2.4 Data Energy 2.5 Implementation

(a) (b) (a) (b) Figure 4: a, W (x); b, dW / dx (x) (a) (b) Figure 1: a, A contour of an object with its 3 features; b, Final deformation with feature attraction; c, Final deformation with data deformation. (a) (b) (c) (d) (e) (f) (a) (b) (c) (g) (h) (a) (c) (d) (e) (b) (a) (b) (c) Future work concentrates on the use of the techniques presented in this paper in comprehensive robotics systems. We have already used deformable surfaces in a system for manipulating objects in natural environments using a three-finger gripper. We are considering using this technique in other systems as well. One candidate is the segmenta-tion of outdoor scenes and the representation of landmarks using a long-range laser range finder 3. The generality of our approach makes it particularly well-suited for unstructured environment in which there are few constraints on object shapes. The scene we use in this example consists in a set of pebbles that are lying on top of sand, using three range images from three different viewpoints (Figure 6(a)). The features (Figure 6(b)) are of three natures: distance discontinuities, surface orientation discontinuities and shadow boundaries. Shadows are parts of the scene that are visible from the camera but are not illuminated by the projector. Final shaded display is shown in Figure 6(c). This application to the segmentation of natural scenes is part of an operational system that automatically picks up pebbles in cluttered natural environment 5. The discrete surface extracted from range images are fed to a program that computes the stable grasp positions for a three-finger gripper 8. To investigate further the ability of the algorithm to perform segmentation and surface reconstruction, we designed a simple scene with three overlapping cubes. Figure 7(a) is a view from one of the two range finders observing the scene: one cube is in balance between two adjacent cubes. The algorithm was applied successively to each of the three cubes. For each segmentation, a region around a cube is selected (square in Figure 7(a)), features are extracted and the center of the initial sphere is set et the center of this region. Figure 7(b)(c)(d) represent the three recovered cubes and Figure 7(e) the final reconstruction of the scene. The algorithm is clearly able to automatically discard features and data that are not consistent with the surface. Among the three reconstructed shapes, the second seems to …