Semi-supervised learning and feature evaluation for RGB-D object recognition

摘要

With new depth sensing technology such as Kinect providing high quality synchronized RGB and depth images (RGB-D data), combining the two distinct views for object recognition has attracted great interest in computer vision and robotics community. Recent methods mostly employ supervised learning methods for this new RGB-D modality based on the two feature sets. However, supervised learning methods always depend on large amount of manually labeled data for training models. To address the problem, this paper proposes a semi-supervised learning method to reduce the dependence on large annotated training sets. The method can effectively learn from relatively plentiful unlabeled data, if powerful feature representations for both the RGB and depth view can be extracted. Thus, a novel and effective feature termed CNN-SPM-RNN is proposed in this paper, and four representative features (KDES [1], CKM [2], HMP [3] and CNN-RNN [4]) are evaluated and compared with ours under the unified semi-supervised learning framework. Finally, we verify our method on three popular and publicly available RGB-D object databases. The experimental results demonstrate that, with only 20% labeled training set, the proposed method can achieve competitive performance compared with the state of the arts on most of the databases.