3D object classification and parameter estimation based on parametric procedural models

Classifying and gathering additional information about an unknown 3D objects is dependent on having a large amount of learning data. We propose to use procedural models as data foundation for this task. In our method we (semi-)automatically define parameters for a procedural model constructed with a modeling tool. Then we use the procedural models to classify an object and also automatically estimate the best parameters. We use a standard convolutional neural network and three different object similarity measures to estimate the best parameters at each degree of detail. We evaluate all steps of our approach using several procedural models and show that we can achieve high classification accuracy and meaningful parameters for unknown objects.