CoarseHash

Benchmark datasets used in ICRA 2020 paper: Fast, Compact and Highly Scalable Visual Place Recognition through Sequence-based Matching of Overloaded Representations [arxiv][IEEE Xplore][YouTube]

Generating Pseudo Localization Datasets

1. Download Ingredient Datasets

Deep1B

The dataset is available here for download. A download script is provided by the original authors.

We only use the descriptors within the base/base00 file, so you may only want to download that single file.

FAS100K

The dataset is described in the paper and its NetVLAD descriptors for both the reference and the query set are available for download here.

The downloaded .npz comprises 7 ndarrays named arr_0 to arr_6, comprising respectively reference descriptors, query descriptors, ground truth match indices for query data, reference poses (xyz), query poses (xyz), ignore, ignore.

2. Generate 20K, 1M, and 10M

Prerequisites

numpy
scikit_learn

See requirements.txt, generated using pipreqs==0.4.10 and python3.5.6

Run

Set the path and dataset variable ("20K", "1M" or "10M") and run python preProcData.py to generate the localization dataset. The "10M" dataset can take around 25 GB of RAM when performing PCA. A low-memory alternative would be to use Incremental PCA.

License

The code is released under MIT License. FAS100K license is as specified on the download link. For Deep1B, refer to its original sources as mentioned above.

If you find this repository useful or use these datasets, cite:

Garg, Sourav, and Michael Milford. "Fast, Compact and Highly Scalable Visual Place Recognition through Sequence-based Matching of Overloaded Representations." In 2020 International Conference on Robotics and Automation (ICRA). IEEE, 2020.

bibtex:

@inproceedings{garg2020fast,
  title={Fast, Compact and Highly Scalable Visual Place Recognition through Sequence-based Matching of Overloaded Representations},
  author={Garg, Sourav and Milford, Michael},
  booktitle={IEEE International Conference on Robotics and Automation (ICRA)},
  year={2020}
}

Related Projects

Delta Descriptors (2020)

seq2single (2019)

LoST (2018)