Handcrafted Features Can Help End-to-End Pose Estimation Using CNN.

This paper addresses the uncertainty and improving the accuracy of end-to-end pose estimators using CNNs, such as in low texture environments. We note that when the number of handcrafted features in the input image of a pose estimation model is small, the model tends to produce output with large errors, and propose a method to reject the model estimate when the number of handcrafted features in the input image is small. The experimental results show that the proposed method improves the accuracy by a factor of up to 7, which demonstrates the usefulness of the proposed method. Our work is important in encouraging the active adoption of deep learning in robot localization, and in enabling simple robots with configurations that can take advantage of a large amount of data.