OCNN: Point Cloud-Based Convolutional Neural Network for Object Orientation Estimation

Convolutional neural networks for pose estimation based on 2D images are severely affected by imaging quality, illumination changes and occlusion, which greatly reduces their reliability and robustness in cluttered scenes. Hence a point cloud based convolutional neural network for orientation estimation(OCNN) is proposed in this paper. With the network, the 3D orientation is estimated by regressing to the rotation matrix. An orthogonality metric loss is introduced to normalize the output rotation matrix. For objects with rotational symmetry, an algorithm for estimating the plane of symmetry and cutting the object is introduced to eliminate orientation ambiguity. Experimental results on LineMOD dataset indicate the accuracy of the angular error below 5° is 94.1%. Combined with advanced point cloud registration algorithms, the network has the potential for real-time orientation estimation.