Recurrent R-CNN: Online Instance Mapping with context correlation

We introduce a novel neural network architecture Recurrent R-CNN using context correlation for online semantic instance segmentation in commodity RGBD scene. Recurrent R-CNN use a new region convolution neural network combined with recurrent unit to assign instance-level semantic class labels to different objects in the scene. We combines with object-level RGB-D SLAM system and makes information to flow in a spatially consistent way across video frames. This dependencies allows significantly higher accuracy and consistency in instance segmentation than state-of-the-art alternatives. Also, we demonstrate a promising augmented reality application under the premise of object-level scene understanding.