Illumination Change Robustness in Direct Visual SLAM

Abstract

Direct visual odometry and Simultaneous Localization and Mapping (SLAM) methods determine camera poses by means of direct image alignment. This optimizes a photometric cost term based on the Lucas-Kanade method. Many recent works use the brightness constancy assumption in the alignment cost formulation and therefore cannot cope with significant illumination changes. Such changes are especially likely to occur for loop closures in SLAM. Alternatives exist which attempt to match images more robustly. In our paper, we perform a systematic evaluation of real-time capable methods. We determine their accuracy and robustness in the context of odometry and of loop closures, both on real images as well as synthetic datasets with simulated lighting changes. We find that for real images, a Census-based method outperforms the others.

Publications

• Illumination Change Robustness in Direct Visual SLAM (ICRA 2017) [pdf] [bib]

Datasets

We base our datasets on the TUM RGB-D benchmark and ICL-NUIM synthetic datasets with depth noise added to synthetic data as done by Choi et al..

Before You Download

We provide our real-world datasets under the Creative Commons CC0 license: To the extent possible under law, the authors have waived all copyright and related or neighboring rights to this work.

For the synthetic datasets based on the ICL-NUIM dataset by A. Handa et al., the original data is used under the CC BY 3.0 license with modifications as described in our paper.

If you use our data, please cite our publication:

@InProceedings{park2017icra,
	author = {Seonwook Park and Thomas Sch\"ops and Marc Pollefeys},
	title = "Illumination Change Robustness in Direct Visual SLAM",
	booktitle = "ICRA",
	year = "2017",
}

Also consider citing the original publication for the ICL-NUIM data.

For using the dataset, please refer to instructions for the TUM RGB-D benchmark. Ground-truth trajectory files can be found at the top-level folder of respective zip archives.

Synthetic Data

These synthetic sequences were generated with tools and trajectories provided by the original ICL-NUIM synthetic datasets with depth noise added via a Python script provided by Redwood.

Real-World Data

These sequences were recorded with a first generation Microsoft Kinect. The format of the dataset is equivalent to that of the TUM RGB-D benchmark (no accelerometer recordings are available). Camera intrinsic parameters used are:

f_x: 538.7 [mm]
f_y: 540.7 [mm]
c_x: 319.2 [pixels]
c_y: 233.6 [pixels]

Depth images are registered via the Kinect firmware.

Acknowledgments

This work was supported by Google and Qualcomm. We thank Torsten Sattler for helpful comments.