How to improve CNN-based 6-DoF camera pose estimation

TL;DR

This paper enhances CNN-based 6-DoF camera pose estimation by exploring dataset characteristics, data augmentation, and LSTM integration to improve accuracy and robustness over existing methods.

Contribution

It introduces modifications such as emphasizing field-of-view, a data augmentation scheme, and LSTM integration to improve PoseNet's accuracy for monocular camera pose estimation.

Findings

Improved pose accuracy with combined modifications.

Field-of-view is more critical than image resolution.

LSTM integration enhances temporal consistency.

Abstract

Convolutional neural networks (CNNs) and transfer learning have recently been used for 6 degrees of freedom (6-DoF) camera pose estimation. While they do not reach the same accuracy as visual SLAM-based approaches and are restricted to a specific environment, they excel in robustness and can be applied even to a single image. In this paper, we study PoseNet [1] and investigate modifications based on datasets' characteristics to improve the accuracy of the pose estimates. In particular, we emphasize the importance of field-of-view over image resolution; we present a data augmentation scheme to reduce overfitting; we study the effect of Long-Short-Term-Memory (LSTM) cells. Lastly, we combine these modifications and improve PoseNet's performance for monocular CNN based camera pose regression.