Evaluation of feature selection and model training strategies for object category recognition

Several methods for object category recognition in RGB-D images have been reported in literature. These methods are typically tested under the same conditions (which we can consider a “domain” in a restricted sense) such as viewing angles, distances to the object as well as lightening conditions on which they are trained. However, in practical applications one often has to deal with previously unseen domains. In this paper, we investigate the effect of domain change on the performance of object category recognition methods. We use the public RGB-D Object Dataset from Lai et al. [1] for training, and for testing we introduce the DLR-RGB-D dataset, representing a similar, but different domain. The data present in both datasets holds various object instances grouped into general object categories. Object category detectors are trained using the objects of one domain and tested on the objects of the other domain. We then explored how do different 3D features perform when the model trained on the source domain is applied on the target domain, and evaluated two feature selection strategies. In our experiments we show that a domain change can have significant impact on the model´s accuracy, and present results for improving the results by increasing the variability of the objects in the training domain. Finally, we discuss the relevance of the descriptors and the properties they capture.