Marco Treiber

Three-Dimensional Object Recognition

Some applications require a position estimate in 3D space (and not just in the 2D image plane), e.g., bin picking applications, where individual objects have to be gripped by a robot from an unordered set of objects. Typically, such applications utilize sensor systems which allow for the generation of 3D data and perform matching in 3D space. Another way to determine the 3D pose of an object is to estimate the projection of the object location in 3D space onto a 2D camera image. There exist methods managing to get by with just a single 2D camera image for the estimation of this 3D → 2D mapping transformation. Some of them shall be presented in this chapter. They are also examples of correspondence-based schemes, as the matching step is performed by establishing correspondences between scene image and model features. However, instead of using just single scene image and model features, correspondences between special configurations of multiple features are established here. First of all, the SCERPO system makes use of feature groupings which are perceived similar from a wide variety of viewpoints. Another method, called relational indexing, uses hash tables to speed up the search. Finally, a system called LEWIS derives so-called invariants from specific feature configurations, which are designed such that their topologies remain stable for differing viewpoints.

Interest Point Detection and Region Descriptors

Object recognition in “real-world” scenes – e.g., detect cars in a street image – often is a challenging application due to the large intra-class variety or the presence of heavy background clutter, occlusion, or varying illumination conditions, etc. These tough demands can be met by a two-stage strategy for the description of the image content: the first step consists of the detection of “interest points ” considered to be characteristic. Subsequently, feature vectors also called “region descriptors ” are derived, each representing the image information available in a local neighborhood around one interest point. Object recognition can then be performed by comparing the region descriptors themselves as well as their locations/spatial configuration to the model database. During the last decade, there has been extensive research on this approach to object recognition and many different alternatives for interest point detectors and region descriptors have been suggested. Some of these alternatives are presented in this chapter. It is completed by showing how region descriptors can be used in the field of scene categorization , where the scene shown in an image has to be classified as a whole, e.g., is it of type “city street,” “indoor room,” or “forest”, etc.?

Flexible Shape Matching

Some objects, e.g., fruits, show considerable intra-class variations of their shape. Whereas algorithms which are based on a rigid object model run into difficulties for deformable objects, specific approaches exist for such types of objects. These methods use parametric curves, which should approximate the object contour as good as possible. The parameters of the curve should offer enough degrees of freedom for accurate approximations. Two types of algorithms are presented in this chapter. Schemes belonging to the first class, like snakes or the contracting curve density algorithm, perform an optimization of the parameter vector of the curve in order to minimize the differences between the curve and the object contour. Second, there exist object classification schemes making use of parametric curves approximating an object. These methods typically calculate a similarity measure or distance metric between arbitrarily shaped curves. Turning functions or the so-called curvature scale space are examples of such measures. The metrics often follow the paradigm of perceptual similarity, i.e., they intend to behave similar to the human vision system.

Geometric Correspondence-Based Approaches

Correspondence-based approaches are often used in industrial applications. Here, the geometry of the objects to be found is usually known prior to recognition, e.g., from CAD data of the object shape. Typically, these methods are feature based, meaning that the object is characterized by a limited set of primitives called features. For example, the silhouette of many industrial objects can be modeled by a set of line segments and circular arcs as features. This implies a two-stage strategy during the recognition phase: in a first step, all possible locations of features are detected. The second step tries to match the found features to the model, which leads to the pose estimation(s) of the found object(s). The search is correspondence based, i.e., one-to-one correspondences between a model feature and a scene image feature are established and evaluated during the matching step. After a short introduction of some feature types, which typically are used in correspondence-based matching, two types of algorithms are presented. The first type represents the correspondences in a graph or tree structure, where the topology is explored during matching. The second method, which is called geometric hashing, aims at speeding up the matching by the usage of hash tables.

Transformation-Search Based Methods

Another way of object representation is to utilize object models consisting of a finite set of points and their position. By the usage of point sets recognition can be performed as follows: First, a point set is extracted from a scene image. Subsequently, the parameters of a transformation which defines a mapping of the model point set to the point set derived from the scene image are estimated. To this end, the so-called transformation space, which comprises the set of all possible transform parameter combinations, is explored. By adopting this strategy occlusion (resulting in missing points in the scene image point set) and background clutter (resulting in additional points in the scene image point set) both lead to a reduction of the percentage of points that can be matched correctly between scene image and the model. Hence, occlusion and clutter can be controlled by the definition of a threshold for the portion of the point sets which has to be matched correctly. After introducing some typical transformations used in object recognition, some examples of algorithms exploring the transformation space including the so-called generalized Hough transform and the Hausdorff distance are presented.