RGB-D Object Recognition Using the Knowledge Transferred from Relevant RGB Images

The availability of depth images provides a new possibility to solve the challenging object recognition problem. However, when there is not enough labeled data, we cannot learn a discriminative classifier even using depth information. To solve this problem, we extend LCCRRD method by kernel trick. First, we construct two RGB classifiers with all labeled RGB images from source and target domain. The significant samples for both classifier are boosted and the non-significant ones are inhibited by exploiting the relationship between two domains. In this process, the knowledge of source RGB classifier can be transferred to target RGB classifier effectively. Then to improve the performance of RGB-D classifier by applying the knowledge from source domain, the predicted results of RGB-D classifier are made consistent to target RGB classifier. Furthermore all the parameters are optimized in a unified objective function. Experiments on four cross-domain dataset pairs shows that our approach is indeed effective and promising.