Probabilistic Regression of Rotations using Quaternion Averaging and a Deep Multi-Headed Network

TL;DR

This paper introduces HydraNet, a deep multi-head network that provides calibrated probabilistic estimates of 3D rotations using quaternion regression and rotation averaging, enhancing motion estimation in AR and robotics.

Contribution

The work extends HydraNet to regress unit quaternions on SO(3) and incorporates uncertainty injection with rotation averaging for improved rotation estimation.

Findings

HydraNet outperforms stochastic methods in uncertainty calibration.

The approach yields consistent orientation estimates on benchmark datasets.

Fusing deep estimates with classical methods improves localization accuracy.

Abstract

Accurate estimates of rotation are crucial to vision-based motion estimation in augmented reality and robotics. In this work, we present a method to extract probabilistic estimates of rotation from deep regression models. First, we build on prior work and argue that a multi-headed network structure we name HydraNet provides better calibrated uncertainty estimates than methods that rely on stochastic forward passes. Second, we extend HydraNet to targets that belong to the rotation group, SO(3), by regressing unit quaternions and using the tools of rotation averaging and uncertainty injection onto the manifold to produce three-dimensional covariances. Finally, we present results and analysis on a synthetic dataset, learn consistent orientation estimates on the 7-Scenes dataset, and show how we can use our learned covariances to fuse deep estimates of relative orientation with classical stereo visual odometry to improve localization on the KITTI dataset.