One the Study of Data Augmentation for Visual Place Recognition

We examined the DA impact per model architectures for VPR from previously unexplored angles. By designing a thorough empirical study, we revealed multiple essential insights:

• The effect of data augmentation is not universally applicable to all VPR models, and even with the same DA, performance may or may not be improved depending on the learning method.
• Even if the dataset, which already contains various realistic constraints, is used for learning, applying data augmentation in classification-based learning methods greatly improves performance.
• Using VPR-inductive biased DA in conjunction with cross-entropy loss makes it easier to optimize VPR performance.

Train

After downloading the SF-XL dataset, simply run

python train.py \
    --data_path PATH/TO/DATASET \
    --backbone ResNet50 \
    --pooling gem \
    --fc_output_dim 4096 \
    --train_aug RRC_RA \
    --use_amp16

Test

You can test a trained model as such

python eval.py \
    --dataset_folder PATH/TO/DATASET \
    --test_queries queries_v2 \
    --backbone ResNet50 \
    --pooling gem \
    --fc_output_dim 4096 \
    --train_aug RRC_RA \
    --resume_model PATH/TO/RESUME

Reproducing the paper

Traning Configuration				Performance
Loss	Pooling	Training Dataset	Augmentation	SF-XL v1	Pitts30k	Tokyo247	Trained model
WSTRL	NetVLAD	Pitts30k	Base	41.0	81.51	62.86	link
			RRC+CJ	37.1	80.34	60.63	link
			RA	40.6	81.32	62.86	link
			RA+CO	41.1	81.28	62.86	link
LMCL	GeM	SF-XL(small)	Base	56.8	81.75	58.73	link
			RRC+CJ	61.7	84.98	72.06	link
			RA	63.5	86.28	69.84	link
			RA+CO	63.4	86.74	70.79	link

Issues

If you questions regarding our code or dataset, feel free to open an issue or send an email to gist.acsl.suji@gmail.com

Acknowledgements

Parts of this repo are inspired by the following repositories:

• CosPlace
• pytorch-NetVlad

Cite

Here is the bibtex to cite our paper