EcL 2 
Mu 
  
  
95 
Pivot element @ Property 1 
Test elements (©) Property 2 
: Sequence T 
Comparisons — q 
ss» =» + 
Pairing > 
mt 
  
Fig.5: Without considering differences in edge 
length incorrect regions are determined for the leg 
area (a). When longer edges are cut in two halves, 
this defect can be remedied. 
  
Fig. 6: Local edge element selection. 
similarity with respect to those properties that are also used to form the multiply linked list mentioned above. 
3. In order to come to a decision for a specific pairing, a sequence of six tests at most is performed on the rela- 
tional attributes of the two edges; the attributes evaluated are: 
— difference in angular direction, 
— positional compatibiliy, excluding collinear cases, 
— proximity, pose, 
— lateral grey value difference, 
— distance, and 
— length deviation. 
If any test fails, subsequent tests will not be performed any more and the pairing under consideration is re- 
jected. In case of compatibility of two edges, the quality of fit is assessed by a scoring function which every test 
contributes to. | 
4. Having finished an evaluation, a pairing is established. If further assignments of a particular edge involved are 
not likely (which has to be guaranteed by various cross checks), this edge will be removed from the set of edges 
for further testing, thus successively reducing search space. 
Compared with a linear optimization approach this algorithm has got the advantage that expensive testing on all 
possible pairings is superfluous. Normally, the association matrix for applications like this one is very sparse. 
The sequence of individual tests and the threshold values for test failure were determined by regression analysis. 
The deviations in attribute values mentioned above have been gathered for more than 10,000 potential pairings in 
different image material (natural, artificial, several object types). Approximately 10% of pairings were interactively 
marked positive, the rest represented failing assignments. On these training data a decision tree was generated 
[Quinlan 86] performing recursive partitioning of the data set like a classifier. As splitting criterion the node im- 
purity function with respect to the selection of the test sequence and the decision threshold was applied. The struc- 
ture of the decision tree gave evidence for a fixed optimal sequence of tests with optimal thresholds. In fig. 7 some 
segmentation results are shown. 
3 Limb identification 
A set of regions given, the next task is deciding for each region what kind of object or subpart it represents or 
whether it belongs to an object that is not related to a human figure. In this matching step, the segmented parts are 
compared with a general appearance model of the articulated body in question. For a first evaluation it is discrimi- 
nated between subparts arm, trunk and legs (fig. 7). A comparison in terms of (sub-) graph isomorphism fails due 
to frequently missing parts in the input image. 
Solving this assignment task is a labeling rather than a classification problem since not the properties of individ- 
ual region segments (primitive attributes) are relevant but rather the geometrical interrelations among regions (re- 
lational attributes). The purpose of labeling is to find a set of consistent limb tags that can explain the image regions 
IAPRS, Vol. 30, Part 5W1, ISPRS Intercommission Workshop “From Pixels to Sequences”, Zurich, March 22-24 1995