Segmentation

Segmentation of a single map deals with assigning labels to the pixel locations of the map so that locations belonging to the same surface patch have the same label. Most segmentation methods for 3-D data are based on the surface geometry and utilize the segmentation techniques developed for conventional gray level images [16].

In [IV], we apply a region-growing algorithm to segment a profile map ${\bf u}_k$ into patches which are planar in the x_k,y_k,z_k frame. The direction of the surface normal is used as a criterion for growing at the first stage. Then, each location on a small segment is added to the neighboring large segment which is closest in the x_k,y_k,z_k frame as measured by the distance from the point to the plane fitted to the data on the neighboring segment. This makes the segmentation more accurate near edges. In [I], we perform only a rough interactive segmentation of a disparity map into quadric patches.

In the case of multiple maps, the question is how to segment the overlapping areas where several maps contain data about the same patch of the object. This question can be addressed in two opposite ways. In the first approach, each map is segmented separately and then the compatible overlapping segments of different maps are merged [22]. In another approach [12], the data sets are first merged and then the whole data are segmented.

We consider the former approach in the case of planar patches in [IV, V]. Our achievement is a split and merge technique for the segmentation of ambiguous areas that may result from the misalignment of the maps. For each pair of maps ${\bf u}_k$ and ${\bf u}_l$, k<l, the segments of ${\bf u}_l$ that overlap several segments of ${\bf u}_k$ are split up in the overlapping area according to the segmentation of ${\bf u}_k$ and the compatible overlapping ones are merged. The algorithm is presented in [IV] and it is further improved in [V] concerning the order in which the segments are processed and using the confidence intervals of the estimated plane parameters to decide on whether two segments are compatible or not. The results of segmentation can be viewed from the reconstructed models in [IV, Fig. 4; V, Fig. 4].