of multiple-view registration. In [43] a global optimization process searches a graph constructed from the pairwise view matches for a connected sub-graph containing only correct matches, using a global consistency measure to eliminate incorrect but locally consistent matches. Other approaches use both global and local prealignment techniques to select the overlapping views by computing a coarse alignment between all the pairs. In [55] the pre-alignment is performed by extracting global features from each view, namely extended Gaussian images. Conversely, in [49], the pre-alignment is computed by comparing the signatures of feature points. Then, the best view sequence is estimated by solving a standard Traveling Salesman Problem (TSP).

6.3.2 Registration in Cluttered Scenes

Thanks to the recent availability of large scale scanners it is possible to acquire scenes composed of several objects. In this context registration is necessary to localize each object present in the scene and estimate its pose. However, in cluttered scenes, an object of interest may be made of a small subset of the entire view. This makes the registration problem more challenging. Figure 6.8 shows two examples of highly cluttered scenes: an entire square2 and a scene composed of several mechanical objects.

Roughly speaking two main strategies have been proposed to address this problem: (i) the use of point signatures to improve point-to-point matching and (ii) the design of more effective matching methods. We now describe each of these in turn.

Point Signatures This approach is similar to local approaches for pre-alignment. Here, due to the cluttered scene, the challenge comes from the fact that the neighborhood of one point of an object can cover part of other objects. Therefore, the descriptor may become useless. In [56] a descriptor that uses two reference points to define a local coordinate system is proposed. In particular, a three-dimensional tensor is built by sampling the space and storing the amount of surface intersecting each sample. In [4] a method that exploits surface scale properties is introduced. The geometric scale variability is encoded in the form of the intrinsic geometric scale of each computed feature, leading to a highly discriminative hierarchical descriptor.

Matching Methods Since the number of corresponding points are very few within cluttered scenes, standard methods for outlier rejection are not useful but more complex matching algorithms can be exploited. In [56], descriptors are stored using a hash table that can be efficiently looked up at the matching phase by a geometric hashing algorithm. In [4], matching is performed in a hierarchical fashion by using the hierarchy induced from the definition of the point-descriptor. In [29], a method

2Piazza Brà, Verona, Italy. Image courtesy of Gexcel: http://www.gexcel.it.