【点云论文速读】PoseNet的改进方法

标题：Geometric loss functions for camera pose regression with deep learning

作者：Alex Kendall and Roberto Cipolla

来源：2017cvpr

星球ID：Lionheart|点云配准

欢迎各位加入免费知识星球，获取PDF论文，欢迎转发朋友圈分享快乐。

●论文摘要

深度学习在单目相机重定位上展示了高效和鲁棒的特性，特别是PoseNet使用深度卷积网络从单个图像中学习相机的六自由度姿态，网络学习使用高级别特征定位特征，并且在不同的光照条件，运动模糊和未知的相机内参的条件下表现鲁棒，在基于SIFT的匹配失效的情况下依旧表现良好，然而，它是使用一个简单的损失函数进行训练的，带有需要昂贵的调优的超参数。本文对这一问题进行了较为基础的理论处理。本文提出了一种基于几何和场景重投影误差的学习相机姿态的新方法。此外，我们还展示了如何自动学习一个最佳加权，以同时倒退的位置和方向。通过利用几何图形，我们证明了我们的技术可以显著提高PoseNet在从室内房间到小城市的数据集上的性能。

●论文图集

●英文摘要

Deep learning has shown to be effective for robust andreal-time monocular image relocalisation. In particular,PoseNet [22] is a deep convolutional neural network whichlearns to regress the 6-DOF camera pose from a single image. It learns to localize using high level features and isrobust to difficult lighting, motion blur and unknown camera intrinsics, where point based SIFT registration fails.However, it was trained using a naive loss function, withhyper-parameters which require expensive tuning. In thispaper, we give the problem a more fundamental theoretical treatment. We explore a number of novel loss functionsfor learning camera pose which are based on geometry andscene reprojection error. Additionally we show how to automatically learn an optimal weighting to simultaneouslyregress position and orientation. By leveraging geometry,we demonstrate that our technique significantly improvesPoseNet’s performance across datasets ranging from indoorrooms to a small city.