Transfering RGB-D Models to RGB by Multi-View Embedding Space Learning for Object Detection

In this paper, we propose a new framework for detecting objects in RGB images captured by conventional cameras by leveraging a set of labeled RGB-D data. We formulate this problem into a new multi-view learning problem, where we have the depth information in the training set that is not available in the testing set. To solve this problem, we use a three-view CCA learning method which combines visual features, depth features, and label information together to get the projection matrices. Sufficient experiments are done to demonstrate the effectiveness of our method, and improvement measures in the future are also proposed.