Why-So-Deep

Official implementation:

• Why-So-Deep: Towards Boosting Previously Trained Models for Visual Place Recognition (RA-L 2022) [Arxiv] | [project website] | [Bibtex]
  • Method (Ours): MAQBOOL: Multiple AcuQitation of perceptiBle regiOns for priOr Learning

Installation

Please refer to INSTALL.md for installation and dataset preparation.

Train & Test

To reproduce the results in papers, you could train and test the models following the instruction in REPRODUCTION.md.

Results

MAQBOOL performance comparison with APANet and NetVLAD at 512-D.

Method	Whitening	Tokyo 24/7			Pitts250k-test
		Recall@1	Recall@5	Recall@10	Recall@1	Recall@5	Recall@10
Sum pooling	PCA whitening	44.76	60.95	70.16	74.13	86.44	90.18
	PCA-pw	52.7	67.3	73.02	75.63	88.01	91.75
NetVLAD	PCA whitening	56.83	73.02	78.41	80.66	90.88	93.06
	PCA-pw	58.73	74.6	80.32	81.95	91.65	93.76
APANet	PCA whitening	61.9	77.78	80.95	82.32	90.92	93.79
	PCA-pw	66.98	80.95	83.81	83.65	92.56	94.7
MAQBOOL (ours)	PCA whitening + DT-50	68.25	79.37	83.49	85.45	92.62	94.58
	PCA whitening + DT-100	69.21	80.32	84.44	85.46	92.77	94.72

If you want to add MAQBOOL results, tested on Pittsburgh and Tokyo247 dataset, in comparison with your work at 4096-D and 512-D. You can use the dat files in the results folder. Names of these dat files are explained in our project page, so that you can easily use. Furthermore, if you need help in plotting the results using Tikz and latex, please follow this little tutorial.

Generate Thumbnails

Thumbnails generating for top 5 results

Open config_wsd.m and change show_output = 1.

Citation

If you find this repo useful for your research, please consider citing the paper

@article{whysodeepBhutta22,
 author={Bhutta, M. Usman Maqbool and Sun, Yuxiang and Lau, Darwin and Liu, Ming},
  journal={IEEE Robotics and Automation Letters}, 
  title={Why-So-Deep: Towards Boosting Previously Trained Models for Visual Place Recognition}, 
  year={2022},
  volume={7},
  number={2},
  pages={1824-1831},
  doi={10.1109/LRA.2022.3142741}}

Acknowledgements

The structure of this repo is inspired by NetVLAD.

Documentation is available at project website. Please follow the installation guide below.

License

Why-So-Deep is released under the MIT license.

⭐️ If you like Why-So-Deep, give it a star here! ⭐️