基于深度学习的平面位姿估计算法

为了在未知物体三维模型的情况下使用深度学习进行平面位姿估计，采用编码器-解码器网络，从单个RGB图像中检测平面实例分割及法线信息，并利用这些信息进行位姿解算，获得每个平面的实时位姿。实验结果显示,平面召回率为0.625，平面法线召回率为0.414，实时性为18.5 f/s，验证了算法的可行性。

Research on Plane Pose Estimation Algorithm Based on Deep Learning

Plane pose estimation is widely used in robotics,augmented reality and other fields,but traditional visual pose estimation algorithms need to be added specific markers to the target,which is less robust and cannot be applied to any scene;pose estimation methods based on deep learning can effectively solve the above-mentioned problems,but the existing method requires a three-dimensional model of a known object.In order to use deep learning for plane pose estimation in the case of unknown object three-dimensional model,the encoder decoder network was used to detect plane instance segmentation and normal information from a single RGB image,which were then employed to perform pose calculation for obtaining each Real-time pose of each plane.The test results show that the plane recall rate is 0.625,the plane normal recall rate is 0.414,and the real-time performance is 18.5 f/s,which verifies the feasibility of the algorithm.