Imitation Learning of Path-Planned Driving Using Disparity-Depth Images

Abstract

Sensor data representation in autonomous driving is a defining factor for the final performance and convergence of End-to-End trained driving systems. When theoretically a network, trained in a perfect way, should be able to abstract the most useful information from camera data depending on the task, practically this is a challenge. Therefore, many approaches explore leveraging human designed intermediate representations as segmented images. We continue work in the field of depth-image based steering angle prediction and compare networks trained purely on either RGB-stereo images or depth-from-stereo (disparity) images. Since no dedicated depth sensor is used, we consider this as a pixel grouping method where pixel are labeled by their stereo disparity instead of relying on human segment annotations.

Cite

Text

Hornauer et al. "Imitation Learning of Path-Planned Driving Using Disparity-Depth Images." European Conference on Computer Vision Workshops, 2018. doi:10.1007/978-3-030-11021-5_33

Markdown

[Hornauer et al. "Imitation Learning of Path-Planned Driving Using Disparity-Depth Images." European Conference on Computer Vision Workshops, 2018.](https://mlanthology.org/eccvw/2018/hornauer2018eccvw-imitation/) doi:10.1007/978-3-030-11021-5_33

BibTeX

@inproceedings{hornauer2018eccvw-imitation,
  title     = {{Imitation Learning of Path-Planned Driving Using Disparity-Depth Images}},
  author    = {Hornauer, Sascha and Zipser, Karl and Yu, Stella X.},
  booktitle = {European Conference on Computer Vision Workshops},
  year      = {2018},
  pages     = {542-548},
  doi       = {10.1007/978-3-030-11021-5_33},
  url       = {https://mlanthology.org/eccvw/2018/hornauer2018eccvw-imitation/}
}