Domain Adaptation for 6-DoF Pose Estimation: Deep Learning with Uncertainty

Abstract

There is considerable research on debris pose estimation using neural networks. However, most researchers have not considered a domain shift, which is a change in the data domain between the training and operational environments. Therefore, we propose a method to estimate the 6-DoF pose in operational environments. The method offers high accuracy even on images obtained during the operation phase by using domain adaptation which reduces domain shift. We have generated simulation and real images that mimic in-orbit images with hardware. These data are labeled. The error rate for tasks in which no domain shift occurs, trained, and estimated on the simulation image, was 5.6%. The error rate for tasks where domain shift occurs, trained on simulation images and estimated real images, was 42.3%. We performed domain adaptation on various dataset with various neural network combinations. Our results in the best combinations showed that unsupervised domain adaptation reduced the error rate to 35.5%, while supervised domain adaptation reduced the error rate to 13.0%. Supervised domain adaptation was able to consider the domain shifts that could not be removed by unsupervised domain adaptation and could estimate pose with high accuracy. Regression problems differs from classification problems in that the effect of small domain shifts directly affects the final output, so even the slight domain shifts cannot be tolerated. For this reason, we found that supervised domain adaptation that can consider domain shift is better suited to pose estimation of the regression problem. In addition, there was no correlation between the degree of uncertainty and estimation error in the unsupervised domain adaptation. However, there was a correlation in supervised domain adaptation.