Reliable Attribute-Based Object Recognition Using High Predictive Value Classifiers

摘要

We consider the problem of object recognition in 3D using an ensemble ofattribute-based classifiers. We propose two new concepts to improveclassification in practical situations, and show their implementation in anapproach implemented for recognition from point-cloud data. First, the viewingconditions can have a strong influence on classification performance. We studythe impact of the distance between the camera and the object and propose anapproach to fuse multiple attribute classifiers, which incorporates distanceinto the decision making. Second, lack of representative training samples oftenmakes it difficult to learn the optimal threshold value for best positive andnegative detection rate. We address this issue, by setting in our attributeclassifiers instead of just one threshold value, two threshold values todistinguish a positive, a negative and an uncertainty class, and we prove thetheoretical correctness of this approach. Empirical studies demonstrate theeffectiveness and feasibility of the proposed concepts.