Application of transfer learning in RGB-D object recognition

In this work, we apply Transfer Learning for a Multimodal Deep learning network for fast and robust object recognition using RGB-D dataset. The ability for a network to train quickly and recognize objects robustly is very important in the field of Robotics. The Multimodal deep learning network avoids time-consuming hand-crafted features and makes use of a RGB-D architecture for robust object recognition. Our architecture has two important features. First, it makes use of both RGB and Depth information of an image to recognize it. To achieve this, our architecture has two CNN processing streams, one for RGB modality and the other for the depth modality. This enables the network to achieve higher accuracy than normal single stream RGB network. We encoded the depth image into colour image before passing it into the CNN stream. The other important feature is the speed of training and improving the accuracy further. To achieve this, we made use of Transfer learning. Firstly we trained a CNN network with 10 classes of different objects and then we transfer the parameters to RGB and depth CNN network. This enables the network to train faster and also achieve higher accuracy for a given number of epochs.