Unsupervised Metric Relocalization Using Transform Consistency Loss

Mike Kasper, Fernando Nobre, Christoffer Heckman, Nima Keivan
Proceedings of the 2020 Conference on Robot Learning, PMLR 155:1736-1745, 2021.

Abstract

Training networks to perform metric relocalization traditionally requires accurate image correspondences. In practice, these are obtained by restricting domain coverage, employing additional sensors, or capturing large multi-view datasets. We instead propose a self-supervised solution, which exploits a key insight: localizing a query image within a map should yield the same absolute pose, regardless of the reference image used for registration. Guided by this intuition, we derive a novel transform consistency loss. Using this loss function, we train a deep neural network to infer dense feature and saliency maps to perform robust metric relocalization in dynamic environments. We evaluate our framework on synthetic and real-world data, showing our approach outperforms other supervised methods when a limited amount of ground-truth information is available.

Cite this Paper


BibTeX
@InProceedings{pmlr-v155-kasper21a, title = {Unsupervised Metric Relocalization Using Transform Consistency Loss}, author = {Kasper, Mike and Nobre, Fernando and Heckman, Christoffer and Keivan, Nima}, booktitle = {Proceedings of the 2020 Conference on Robot Learning}, pages = {1736--1745}, year = {2021}, editor = {Kober, Jens and Ramos, Fabio and Tomlin, Claire}, volume = {155}, series = {Proceedings of Machine Learning Research}, month = {16--18 Nov}, publisher = {PMLR}, pdf = {https://proceedings.mlr.press/v155/kasper21a/kasper21a.pdf}, url = {https://proceedings.mlr.press/v155/kasper21a.html}, abstract = {Training networks to perform metric relocalization traditionally requires accurate image correspondences. In practice, these are obtained by restricting domain coverage, employing additional sensors, or capturing large multi-view datasets. We instead propose a self-supervised solution, which exploits a key insight: localizing a query image within a map should yield the same absolute pose, regardless of the reference image used for registration. Guided by this intuition, we derive a novel transform consistency loss. Using this loss function, we train a deep neural network to infer dense feature and saliency maps to perform robust metric relocalization in dynamic environments. We evaluate our framework on synthetic and real-world data, showing our approach outperforms other supervised methods when a limited amount of ground-truth information is available.} }
Endnote
%0 Conference Paper %T Unsupervised Metric Relocalization Using Transform Consistency Loss %A Mike Kasper %A Fernando Nobre %A Christoffer Heckman %A Nima Keivan %B Proceedings of the 2020 Conference on Robot Learning %C Proceedings of Machine Learning Research %D 2021 %E Jens Kober %E Fabio Ramos %E Claire Tomlin %F pmlr-v155-kasper21a %I PMLR %P 1736--1745 %U https://proceedings.mlr.press/v155/kasper21a.html %V 155 %X Training networks to perform metric relocalization traditionally requires accurate image correspondences. In practice, these are obtained by restricting domain coverage, employing additional sensors, or capturing large multi-view datasets. We instead propose a self-supervised solution, which exploits a key insight: localizing a query image within a map should yield the same absolute pose, regardless of the reference image used for registration. Guided by this intuition, we derive a novel transform consistency loss. Using this loss function, we train a deep neural network to infer dense feature and saliency maps to perform robust metric relocalization in dynamic environments. We evaluate our framework on synthetic and real-world data, showing our approach outperforms other supervised methods when a limited amount of ground-truth information is available.
APA
Kasper, M., Nobre, F., Heckman, C. & Keivan, N.. (2021). Unsupervised Metric Relocalization Using Transform Consistency Loss. Proceedings of the 2020 Conference on Robot Learning, in Proceedings of Machine Learning Research 155:1736-1745 Available from https://proceedings.mlr.press/v155/kasper21a.html.

Related Material