An Evolutionary Approach for Learning Motion Class Patterns

Abstract

Abstract. This article presents a genetic learning algorithm to derive discrete patterns that can be used for classification and retrieval of 3D motion capture data. Based on boolean motion features, the idea is to learn motion class patterns in an evolutionary process with the objective to discriminate a given set of positive from a given set of negative training motions. Here, the fitness of a pattern is measured with respect to precision and recall in a retrieval scenario, where the pattern is used as a motion query. Our experiments show that motion class patterns can automate query specification without loss of retrieval quality.

Generating Concise Rules for Human Motion Retrieval

“…The retrieval performance of our method is compared with the existing heuristic method [8] under almost the same experimental condition. We experimentally retrieved motion segments from a large public collection of motion capture data, called Mocap Database HDM05 (http://www.mpiinf.mpg.de/resources/HDM05/).…”

“…The training dataset consists of 7 motion classes as summarized in Table 2. The number of training motion classes is fewer than [8] because ILP is suited for the learning of a small number of classes with minimal training dataset. The retrieval rule of each class is discovered by analyzing the training motions of one class and the other six as the positive and negative examples, respectively.…”

“…However, the representation of a motion template is often redundant in attempting to classify the motions because it is simply created by the weighted average of multiple motion segments of the same target class. Our method is similar to the heuristic method which eliminates the redundancy of the motion template using a genetic algorithm [8], [9]. This method reduces the redundant components of the motion template using a genetic algorithm.…”

“…Furthermore, the template cannot be arbitrarily edited because it is difficult to determine which template component is important for the similarity search. Although the heuristic method [8], [9] eliminates such redundancy, the conventional method does not guarantee the minimality of the resulting template and requires high computational cost.…”

Texture Classification of Proteins Using Support Vector Machines and Bio-inspired Metaheuristics

“…In (1) the number of genes with a true binary value (feature selected) is represented by numberActiveFeatures. Regarding classification results, taking into account the F-measure apparently gives better results than only using the accuracy obtained with image features [44,45]. F-measure (2) is a function made up of the recall (true positives rate or sensitivity: proportion of actual positives which are correctly identified) and precision (or positive predictive value: proportion of positive test results that are true positives) measurements.…”