Next: Relaxation Labelling Methods Up: Model Based Object Recognition Previous: Tree Search Methods

### Some Matching Case Studies

Basically many constraints have been developed in order to produce efficient search heuristics. Here we describe some popular methods which illustrate a broad range of heuristics.

The Oshima and Shirai Method (1979)

This method was one of the first methods developed to deal with a variety of classes of objects and to be able to recognise a scene described by depth maps containing multiple objects in any orientation. Objects are considered to have planar surfaces, smoothly curved surfaces or both. The classes of smoothly curved surfaces allowed include ellipsoids, hyperboloids, cones, paraboloids and cylinders.

This method operates in two stages.

The Learning Phase
This method does not use any stored solid model representation to perform matching. Instead models are constructed from images of the objects to be recognised. Thus, a learning phase takes place in which known objects are individually shown to the system in different poses. A description of each object is built up, in terms of properties of surfaces and their relationships, to create a model of the object.

The description is made using the segmentation of curved and planar surfaces from individual images. The segmentation method employed by Oshima and Shirai is a basic region growing method except that many surface properties and interrelationships are also calculated. These then are represented in a relational graph or semantic network. Typical surface properties for each region include:

• surface type (plane, cylinder etc.),
• area,
• perimeter,
• minimum, maximum and mean radii from the surface centroid to the boundary,and
• the standard deviation of the radii from the surface centroid to the boundary.

The relationships between surfaces are characterised by:

• distance between surface centroids,
• angle between best fit planes to this and neighbouring regions (if the surface is curved the best fit plane is still used for this measure),
• type of intersection -- classed as convex, concave, mixed (if intersection is not concave or convex) or no intersection.

Only a discrete set of views are used to model an object. Thus, for a complete model description of an object to be compiled, all unique views of the object must be presented to the system.

The Matching Phase
This phase operates by trying to match the set of visible regions observed in a scene to each object view in a database of learned object descriptions. Not all surfaces present in the corresponding model view of the scene may be present in the observed scene description because of occlusion or poor segmentation, so a procedure that restricts incorrect matchings is required. This is achieved by initially driving the search with observed data and then, when a suitable set of scene features have been successfully matched with the model, model data is used to efficiently guide the rest of the search (see Fig. 56

Fig.  Oshima and Shirai matching method The image data driven part of the search involves constructing one or more kernels  , each of which contains one or two regions that provide the best potential matches to an object model. The chosen regions are selected according to the criteria that they should, if possible, be planar, have large area, and have many neighbours. The criteria are relaxed until ideally two regions or at worst one region for the kernel is found.

The model database is then searched for matches starting with the best kernel. If no match is found then another kernel is selected. If only one match is found using this kernel then the model driven matching phase is invoked. If more than one match is found then other possible kernels are also considered. The system suspends any further matching for kernels which give more than one match until all other kernels providing a single match have been considered by the model driven phase. Note that at the end of the matching process there could be more than one interpretation for a given kernel.

The model driven matching process of the second phase involves searching the scene around the kernel for regions corresponding to those in the model. The search is controlled by a similar type of depth first tree search that which we have already described.

At all stages regions are considered to be matched if a similarity function , based on region properties and relationships, is below a certain threshold.

When a match for a particular kernel has been found, a candidate body is said to have been found in the scene. The matching process continues until all scene regions have been tested for participating in a match. If only one candidate body is found in the scene then it is accepted as the desired match. Otherwise the scene requires further interpretation. If there are multiple candidates, they are examined to filter out inconsistent interpretations: two candidate bodies must be inconsistent if they share a region. For each kernel found a candidate body is rejected if one or more of its features are included in another candidate body which is consistent with respect to the same kernel. At the end of this process, hopefully one or perhaps more sets of mutually consistent interpretations will remain. Each such combination gives a possible interpretation of the scene from which position and orientation information can be calculated.

The 3DPO Method (1983)

The 3DPO (Three-dimensional Part Orientation) vision system was developed by Bolles and his colleagues to recognise and locate parts from depth maps for the purposes of directing a robot arm to grasp a wide range of moderately complex industrial parts. To achieve this goal they argued that most current vision strategies did not provide the required generality, being based upon simple generic object types such as polyhedra, cylinders, spheres or surface patches. Instead they adopted an entirely different approach that could recognise objects using only a few features or feature clusters that highly constrain the object (such as a circle with a fixed radius) and consequently the size of the solution search space. These ideas are now known as local feature focus methods.

The matching process consists of five stages:

Primitive Feature Detection
-- Here low-level features are segmented and linked together. Edges are detected in the three-dimensional data. Edges are classified as being convex, concave or step edges.
Feature Cluster Formation
-- The aim of this stage is to form clusters of (primitive) features that can be used to hypothesise the type, position and orientation of the object. Thus coplanar edge clusters are formed and circular arcs are separated out, for example.
Hypothesis Generation
-- A search is made of the database of models to find a match. The strategy used here is to test individual clusters first and if these are not sufficient then groups of clusters are tested. It is at this stage that the feature focus approach is used by searching for object specific clusters such as circles of a certain radius.
Hypothesis Verification
-- Next, any hypotheses formed are tested by looking for additional features in the image that are consistent with each given hypothesis.
Parameter Refinement
-- In this final stage a given verified hypothesis is interrogated again in order to produce a more accurate location for the part. Averaging or least squares techniques are applied during this step.

This matching process uses an extended solid model. Here a form of boundary model) is augmented with a feature classification network. The feature network classifies features according to type and size, and groups features that share common surface normals, common axes (for cylinders) etc.

The ACRONYM Method (1979)

Brooks has developed a vision system, ACRONYM, that recognises three-dimensional objects occurring in two dimensional images. The examples given in the papers involve recognising airplanes on the runways of an airport from aerial photographs. He uses a generalised cylinder approach to represent both stored model and objects extracted from the image. . As in Nevatia and Binford's method (See next section), a relational graph structure is used to store the representation. Nodes are the generalised cylinders and the links represent the relative transformations (rotations and translations) between pairs of cylinders.

The system also uses two other graph structures, which are constructed from the object models, to help in the matching strategy:

Restriction Graph
-- This graph represents constraints on classes of objects, in order to make generic part descriptions more specific. An example given by Brooks considers classes of electric motors. These can be described by a generic motor type which is then divided into more specific classes of motors such as motor with a base, motor with flanges. These can then be further described in terms of functional classes (dependent on use) such as central heating water pump or gas pump. Additional constraints are allowed to be added to the graph during the recognition process.
Prediction Graph
-- Links in this graph represents relationships between features in the image. The links are labelled must-be, should-be or exclusive according to how likely it is that a given pair of features will occur together in a single object.

Since the image is only two-dimensional account must be taken of the projection of the three-dimensional object features into two dimensions. Brooks uses two descriptions for two-dimensional features:

Ribbons
-- These are used to describe a projection of a body made of generalised cylinders. Note that Nevatia and Binford use a three-dimensional form of ribbon.
Ellipses
-- These are used to describe the projection of the ends of a generalised cylinder. For ends of a circular cylinder the projections are exactly ellipses. For polygonal cross-sections, ellipses can provide a description of the ends by fitting the best circle or ellipse through the vertices and noting the projection of this shape.

The matching process is performed in two stages:

1. The image is searched for local matches to ribbons derived from the stored model. Such instances of ribbon matches are grouped into clusters.
2. The clusters are checked for global consistency in that each match must satisfy both the constraints of the prediction graph, and the accumulated constraints of the restriction graph.

This work has been extended by Kuan and Drazowich to permit use of three-dimensional images. The basic principles of ACRONYM are adhered to but surface properties are incorporated. Also, numerical feature measurements are used during matching (whereas there is a purely symbolic use of features in the original method). The surface properties are included in a multilevel representation of objects which is used in the prediction phase. The levels and the features represented at each level are:

Object Level
-- stores, for example, spatial relationships between object components, object dimensions, view point characteristics such as minimum and maximum size, and occlusion relationships between features.
Generalised cylinder level
-- stores cylinder contour, cylinder position and orientation, cylinder dimensions, relationship between edges etc. This level is very important since generalised cylinders are the basic representation of objects in the system. Specific cylinder properties are therefore kept separate from more general surface characteristics.
Surface Level
-- describes surface type, edge boundary information, and spatial surface relationship.,
Edge Level
-- records edge types (step, convex, concave etc.).

The Grimson and Lozano-Perez Method (1984)

This method was originally designed to match sparse oriented surface points (points on an object with a surface normal attached) but has also been extended to match straight edges by Murray (1988) and others. Here we discuss the method in terms of matching straight line edges in a scene to straight line edges of a model, based on the work by Murray, as edges are more suitable than points when dense depth data is available. However, the discussion is equally valid for the case of oriented surface points. The only differences are that the points are matched to planar model faces and the implementation of some geometric constraints is slightly different.

The method requires the following assumptions:

• The image data have been segmented to provide edges that are accurate to within some specified error. This error is specified in terms of the positioning of the end points and an uncertainty cone for the direction of the edge. Oriented surface point errors are specified in terms of a small volume of error.
• The objects to be recognised can be modelled as polyhedra.

This method is divided into two stages:

1. An exhaustive set of feasible interpretations of the sensed data is produced. Interpretations consist of pairings of each sensed edge with some stored model edge. Interpretations inconsistent with local edge constraints derived from the model are simply discarded. The edge constraints are said to be local since they are derived from knowledge of neighbouring edges only and not from overall knowledge gained from the whole view or object model.
2. Each feasible interpretation is tested for consistency with the object model. An interpretation is allowable if it is possible to provide a rotation and translation that places each sensed edge in correspondence with a model edge to within some specified error.

The two stages are now described separately.

Generating Feasible Interpretations

If an object model contains m edges and there are s sensed edges in the scene then there are possible sets of complete interpretations which match each sensed edge with some model edge. Only a few interpretations lead to satisfactory solutions. Even for simple three-dimensional objects the number of possible interpretations is very large. For example, a view of a cube can show up to nine out of a possible twelve edges giving possible interpretations of which 24 are plausible. It should be noted that this number can be very much larger greater still when sampled oriented points are employed instead of edges. Clearly it is not possible to exhaustively test all these interpretations for a match so the basic idea is to exploit some simple, local geometric constraints on the sensed edges to minimise the number of interpretations that must be tested explicitly.

The possible pairings of sensed and model edges are represented in the form of a search tree, called an interpretation tree.

At each level of the tree, one of the edges from the scene is matched with each of the m possible edges in the model. Each node has m children representing taking the match proposed so far together with all possible matches for the current scene edge. The result is a tree where complete paths to the deepest level correspond to complete matchings of object and scene edges. In the example shown in Fig. 55, we assume (unrealistically) for the purposes of illustration that the model has 4 edges and 3 edges have been found in the scene. Thus the pairings ((1,2),(2,1),(3,1)) corresponds to edge 1 in the scene matching edge 2 in the model, edge 2 in the scene matching edge 1 in the model and edge 3 in the scene matching edge 1 in the model. Note that because of errors in segmentation and other causes such as partial occlusion, it may be possible for more than one edge in the scene to correctly match a single edge in the model.

The interpretation tree is searched in depth first manner and pruned by rejecting interpretations that fail to meet one or more local geometric constraints. If the test is unsuccessful the current node and all of its descendants may be rejected as none of these descendants can possibly correct the current mismatch. In this way the potentially large search space is controlled.

The three local geometric constraints employed to prune the interpretation tree are:

Distance Constraint
-- The length of the sensed edge must be less than or equal to the length of the model edge under consideration.
Angle Constraint
-- The angle between two adjacent sensed edges must agree with that between the two corresponding matched model edges.
Direction Constraint
-- Let represent the range of vectors from any point on sensed edge a to any point on sensed edge b. In an interpretation which respectively pairs sensed edges a and b with model edges i and j, this range of vectors must be compatible with the range of vectors produced by i and j.

The constraints used if matching oriented surface points to planar model faces are simpler:

Distance Constraint
-- The distance between each pair of sensed points must be a possible distance between the corresponding faces.
Angle Constraint
-- The angle between each pair of oriented sensed points must be compatible with that between the known surface normals of the corresponding faces.
Direction Constraint
-- The vector linking two sensed points must lie in a direction which lies within the range of all possible vectors going between the respective surfaces they belong to.

An advantage of this constraint based approach is that all the model information required by the constraints can be computed off-line and stored in look up tables. Any information that is required for comparisons is then simply referred to as needed.

Model Testing

This second stage involves interrogating the sets of possible interpretations which are left after the first stage in order to firstly, estimate the transformation (rotation, scale and translation) between scene and model coordinates, and secondly, to check that all edges or surface points under the estimated transformation lie within acceptable bounds of the assigned edges or assigned faces.

The transformation is estimated by fitting the scene's description to the matched model edges. Two approaches can be applied to solve the problem. The first approach involves simply averaging the estimates of the transformation obtained for each matching pair. Details of the mathematics involved in this approach can be found in Grimson and Lozano-Perez's original paper or the Marshall and Martin Book. Alternatively, to produce more stable results, an approach similar to the estimation of rotation and translation presented in Marshall and Martin Book may be employed. This method is due to Faugeras and Hebert and has been employed by Murray and in the TINA vision system (discussed shortly).

Related Grimson and Lozano-Perez Methods

Apart from the Murray matching algorithm, several other approaches have been developed over recent years based on the Grimson and Lozano-Perez method. In particular Sheffield University's AI Vision Research Unit (AIVRU) have extended this strategy for their TINA vision system. This will be described in the next Section.

Lozano-Perez and Grimson have also developed their strategy to take into account overlapping and mismatching faces. This is achieved by incorporating an extra branch at each node in the interpretation tree to allow for null matches. This technique is described in detail later in Marshall and Martin Book as well as their papers..

Ikeuchi has developed a method for object manipulation tasks. He generates an interpretation tree from object views and then proceeds to orientate and locate objects in the scene using extended Gaussian images.

Recently, Brady and his colleagues have proposed a method of recognising objects described by parameterised models. They aim to allow recognition of objects that do not have or do not require explicit or fixed representations. The example cited in the paper is that of a vision system mounted on an autonomous guided vehicle navigating around a factory environment. This system must recognise wooden pallets used for stacking. A pallet can have many valid forms which humans readily recognise. However, a system using fixed object models would only be able to recognise an observed pallet if it were lucky enough to have a model of that particular pallet. To overcome this a parameterised object model is used where each wooden slat's position relative to some reference point is allowed to vary in the model, as shown in Fig. 57. The value of each parameter is determined in the matching process along with the rotation and orientation.

Fig.  Parameterised representation of a pallet The TINA Method

Sheffield University's AI Vision Research Unit (AIVRU) have developed a matching strategy for their TINA vision system. Much of their work has recently been published as a book containing a collection of papers related to the TINA vision system and other important aspects of three-dimensional model-based matching. Their method uses an approach similar to that of Grimson and Lozano-Perez except that the depth-first tree search is replaced by a hypothesise and test strategy. This grows feasible interpretations in order to improve the efficiency of the search. When some information in a scene may be missing (for example, when occlusion  of an object occurs) tree searches tend to be combinatorially explosive. The hypothesise and test strategy is driven by selecting features from the image based on feature sizes and neighbourhoods. This hypothesise and test strategy is similar to the feature focus} methods used in the 3DPO matching algorithm  discussed in for 3DPO.

The system matches three-dimensional edges (obtained by the use of passive stereo) to three-dimensional edge based models. The matching algorithm uses similar pairwise local geometric relationships to those of Grimson and Lozano-Perez and Murray to limit the search space further.

The basic method is as follows:

1. A edge is chosen from the model to focus the search on. The edge is determined on the basis of its length and the relative length of other edges within a predefined distance from it.
2. A set of edges close to the focus edge that exceed some predefined length is determined.
3. Potential matches to the focus edge are identified in the image on the basis of their length.
4. This set of potential matches is narrowed down by testing for consistency with neighbouring edges.
5. This set of matches is searched for maximally consistent subsets that can be represented by a single matching transformation (rotation and translation). The subsets must contain at least a certain minimum number of elements. Subsets that represent the same transformation are merged.
6. Each subset is extended by adding new matches for all edges in the scene which are consistent with the current set of matches.
7. The extended subsets are ranked on the basis of number and length of their elements.
8. A final transformation is recovered using the least squares method of Faugeras and Hebert mentioned above.

The Faugeras and Hebert Method (1983)

This method matches primitive surfaces extracted from a scene to ones in the model. In the original approach the method is described as coping with planar and quadric surfaces.

The basic approach is to try to match each primitive surface from the model, say , to one segmented from the scene, , to obtain a match pair . This is achieved using a tree search similar to the interpretation tree described earlier. However, an estimation of the transformation (rotation and translation) describing the mapping from model to scene is used as a constraint to control the search as it proceeds, in contrast to the previous methods based on the Grimson and Lozano-Perez method where the transformation is only estimated after matching is completed.

From a small initial set of match pairs, the match list, an initial estimate of the transformation is made. Pairs of primitives from the scene and model are then added to this list if they are compatible with the match so far. Each time a successful match is found a new estimate of the transformation is calculated, and the process is continued until all scene primitives have been matched. If at any stage no possible match pair can be found then a backtracking mechanism is used. A match is considered to be successful if it is consistent with the existing transformation within an allowed tolerance.

Thus, the method exploits the global constraint of rigidity under the transformation to control the potentially combinatorially explosive search space. This is a major difference from the Grimson and Lozano-Perez method. The resulting paradigm of the Faugeras and Hebert method is that of recognising whilst locating, which is controlled mainly by global geometric constraints and not by testing a set of locally generated feasible interpretations. The use of global geometric constraints to control the search is generally more efficient.

At each step the estimation of the transformation is found by solving a least squares minimisation problem (not considered here -- details Marshall and Martin Book or original papers).

A cheap constraint is also employed at each level of the search in order to eliminate grossly mismatching pairs more quickly. The purpose of these tests is to provide a rough estimate of the geometric consistency of the new pair with the current match list. This test can be either or both of:

1. The angle between a matched primitive in the scene and the current scene primitive is equal to the angle between the corresponding primitives in the model to within a given tolerance. Any previously matched scene primitive can be used in principle to implement the constraint although usually one of the first matched primitives is employed since we can order scene primitives before the matching process to make the matching process more efficient. This can be done by choosing the best segmented features or the features most likely to define the matching transformation completely early on in the search or both.
2. The angle between the current scene primitive and the current model primitive rotated by the current estimation of the rotation is equal to zero, to within a given tolerance.

This method has been used in many different computer vision applications. It has been employed in matching in the two-dimensional domain, matching different depth maps and matching using simple occluding (or jump) edges and surface primitives. In the last case Hebert and Kanade argue that because segmentation of high order surface patches is unreliable, low level primitives should be used first to obtain a small number of possible solutions which are then tested using a detailed surface matcher. They employ simple occluding edges which are easily and reliably calculated to limit the number of possible interpretations.

Next: Relaxation Labelling Methods Up: Model Based Object Recognition Previous: Tree Search Methods

dave@cs.cf.ac.uk