CAMAL: Context-Aware Multi-scale Attention framework for Lightweight Visual Place Recognition

• Three benchmark datasets tested on the proposed methodology:

1. SPEDTest
2. St Lucia
3. Synthesized Nordland

If you use the datasets/results, please cite the following publication:

@article{khaliq2019camal,
  title={CAMAL: Context-Aware Multi-scale Attention framework for Lightweight Visual Place Recognition},
  author={Khaliq, Ahmad and Ehsan, Shoaib and Milford, Michael and McDonald-Maier, Klaus},
  journal={arXiv preprint arXiv:1909.08153},
  year={2019}
}

• Each dataset contains two traverses of the same route under different conditions and viewpoint

  • A "VPR_Results" folder in each dataset contains CSV files of all the evaluated VPR technqiues
    • All the CSV files have same pattern; each row contains four paramters i.e. (Test Image number, Retrieved Image number, Score, Matched(1/0)?)
    • The user needs to use the Score and Matched values for drawing the PR-curves

• Another "Vocabulary" folder contains N=300 and V=128 clustered regional dictionary trained using 3K images for VLAD retrieval.

• A python script "produceResults.py" can generate the AUC-PR.

Configuration : N = 300, V=128 (AUC-PR Results)

1. SPEDTest: 0.837
2. St Lucia: 0.747
3. Synthesized Nordland: 0.737