Multi-stage feature learning based object recognition and 3D pose estimation with kinect

Impressive achievements have been achieved over the years in object recognition and 3D pose estimation while detecting objects and estimating their 3D poses reliably is still a very challenging problem, especially for texture-less objects in heavily cluttered scenes. In this paper, we introduce an improved approach based on multi-stage feature learning to identify object as well as their 3D orientation efficiently. Our approach is motivated by visual neuroscience models. Currently similar architectures have been successfully applied to pedestrian detection, traffic recognition, face representation, etc. We compute the feature descriptors by multi-stage convolutional neural networks (CNNs) which could map the object identity and 3D poses to the Euclidean distance space by evaluating the similarity between descriptors. By contrast with classical approaches, our method relies on fast and approximate Nearest Neighbor (NN) matching to solve the large scale problems instead of one classifier per object or multi-class classifiers whose complexity grows rapidly with the number of objects. The experiments of detecting poorly-textured objects in cluttered scenarios have been carried out to verify the effectiveness and efficacy of the proposed approach.