Use of LSTM Regression and Rotation Classification to Improve Camera Pose Localization Estimation

摘要

More accurately estimating camera pose can be used to greatly improve localization in applications such as augmented reality, autonomous driving, and intelligent robots. Deep learning methods have achieved great progress to improve accuracy but still have limitations with respect to rotation, which results in angle regression errors. In this work, we combine a LSTM module with rotation classification loss to regress the camera pose. The algorithm uses a robust processing pipeline to supervise the pose estimation with dynamic, weighted, multi-losses in order to limit separate Euler angle (yaw, pitch, roll) losses, and common translation-quaternion losses. An empirical test on the 7Scenes benchmark dataset shows better results than when using common absolute pose regression methods.