Trend Shift in Pose Estimation: Toward Continuous Representation

Pose estimation is a fundamental task in many applications that utilizing 3D data from sensors like LiDAR and RGB-D cameras. It is particularly crucial in fields where precise position and orientation information are required, such as autonomous driving, cooperative perception, robotics, and augmented reality (AR). To improve the pose estimation, many methods used in other applications are adopted to enhance the network architecture and these methods make significant progress in pose estimation. However, when using deep neural networks (DNNs), the issue of discontinuous rotation representation has emerged, and various studies pointed out that this could be a cause of substantial error. Therefore, we focus on addressing the issue of discontinuities by reviewing the latest research trends in pose estimation published by major academic publishers and provide insights into future directions for pose estimation.