Classifier combination for sketch-based 3D part retrieval

In this paper, we present a search method with multi-class probability estimates for sketch-based 3D engineering part retrieval. The purpose of using probabilistic output from classification is to support high-quality part retrieval by motivating user relevance feedback from a ranked list of top categorical choices. Given a free-hand user sketch, we use an ensemble of classifiers to estimate the likelihood of the sketch belonging to each predefined category by exploring the strengths of various individual classifiers. Complementary shape descriptors are used to generate classifiers with probabilistic output using support vector machines (SVM). A weighted linear combination rule, called adapted minimum classification error (AMCE), is developed to concurrently minimize the classification errors and the log likelihood errors. Experiments are conducted using our Engineering Shape Benchmark database to evaluate the proposed combination rule. User studies show that users can easily identify the desired classes and then the parts under the proposed method and algorithms. Compared with the best individual classifier, the classification accuracy using AMCE increased by 7% for 3D models, and the average best rank improved by 11.6% for sketches.