CollectionsCodeDatasets

seq2single

oravus/seq2single

Look No Deeper: Recognizing Places from Opposing Viewpoints under Varying Scene Appearance using Single-View Depth Estimation

This is the source code for the paper titled: "Look No Deeper: Recognizing Places from Opposing Viewpoints under Varying Scene Appearance using Single-View Depth Estimation", [arXiv][IEEE Xplore].

If you find this work useful, please cite it as: Garg, S., Babu V, M., Dharmasiri, T., Hausler, S., Suenderhauf, N., Kumar, S., Drummond, T., & Milford, M. (2019). Look no deeper: Recognizing places from opposing viewpoints under varying scene appearance using single-view depth estimation. In IEEE International Conference on Robotics and Automation (ICRA), 2019. IEEE.

bibtex:

@inproceedings{garg2019look,
title={Look No Deeper: Recognizing Places from Opposing Viewpoints under Varying Scene Appearance using Single-View Depth Estimation},
author={Garg, Sourav and Babu V, Madhu and Dharmasiri, Thanuja and Hausler, Stephen and Suenderhauf, Niko and Kumar, Swagat and Drummond, Tom and Milford, Michael},
booktitle={IEEE International Conference on Robotics and Automation (ICRA)},
year={2019}
}

Illustration of the proposed approach

An image depicting topometric representation.

Requirements

  • Ubuntu (Tested on 16.04)
  • Jupyter (Tested on 4.4.0)
  • Python (Tested on 3.5.6)
    • numpy (Tested on 1.15.2)
    • scipy (Tested on 1.1.0)

Optionally, for vis_results.ipynb:

  • Matplotlib (Tested on 2.0.2)

Download an example dataset and its pre-computed representations

  1. In seq2single/precomputed/, download pre-computed representations (~10 GB). Please refer to the seq2single/precomputed/readme.md for instructions on how to compute these representations.

  2. [Optional] In seq2single/images/, download images (~1 GB). These images are a subset of two different traverses from the Oxford Robotcar dataset.

(Note: These download links from Mega.nz require you to first create an account (free))

Run

  1. The Jupyter notebook seq2single.ipynb first loads the pre-computed global image descriptors to find top matches. These matches are re-ranked with the proposed method using the pre-computed depth masks and dense conv5 features.

License

The code is released under MIT License.

Related Projects

Delta Descriptors (2020)

CoarseHash (2020)

LoST (2018)

CRICOS No. 00213J