Visual place recognition is a challenging task in computer vision and a key component of camera-based localization and navigation systems. Recently, Convolutional Neural Networks (CNNs) achieved high results and good generalization capabilities. They are usually trained using pairs or triplets of images labeled as either similar or dissimilar, in a binary fashion. In practice, the similarity between two images is not binary, but rather continuous. Furthermore, training these CNNs is computationally complex and involves costly pair and triplet mining strategies. We propose a Generalized Contrastive loss (GCL) function that relies on image similarity as a continuous measure, and use it to train a siamese CNN. Furthermore, we propose three techniques for automatic annotation of image pairs with labels indicating their degree of similarity, and deploy them to re-annotate the MSLS, TB-Places, and 7Scenes datasets. We demonstrate that siamese CNNs trained using the GCL function and the improved annotations consistently outperform their binary counterparts. Our models trained on MSLS outperform the state-of-the-art methods, including NetVLAD, and generalize well on the Pittsburgh, TokyoTM and Tokyo 24/7 datasets. Furthermore, training a siamese network using the GCL function does not require any pair mining.

The code is licensed under the MIT License.

If you use our code please cite our paper

@article{leyvavallina2021gcl,
  title={Data-efficient Large Scale Place Recognition with Graded Similarity Supervision}, 
  author={María Leyva-Vallina and Nicola Strisciuglio and Nicolai Petkov},
  journal={CVPR},
  year={2023}
}

If you have any doubts please contact us at:

1. María Leyva-Vallina: m.leyva.vallina at rug dot nl
2. Nicola Strisciuglio: n.strisciuglio at utwente dot nl

1. MSLS: The dataset is available on request here. For the new GT annotations, please request them here.
2. Pittsburgh: The whole dataset is available on request here and the train val splits for Pitts30k are available here.
3. TokyoTM: The dataset is available on request here.
4. Tokyo 24/7: The dataset is available on request here.
5. TB-Places: The dataset is available here. For the new GT annotations, please request them here.

All our models (and labels) can be downloaded from here.

Run the src/labeling/create_json_idx.py file to generate the necessary json index files for the dataset.

Clone the mapillary repository and run the following command on your machine, substituting "MYDIR" with the path where you downloaded the mapillary library:

export MAPILLARY_ROOT="/MYDIR/mapillary_sls/"

Create the JSON index files for the MSLS dataset as follows. (replace PATH-TO-DATASET with the directory where the MSLS dataset is)

python3 src/labeling/create_json_idx.py --dataset msls --root_dir PATH-TO-DATASET/MSLS/

Run the extract_predictions.py script to compute the map and query features, and the top-k prediction. For instance:

python3 extract_predictions.py --dataset MSLS --root_dir PATH-TO-DATASET/MSLS/ --subset val --model_file models/MSLS/MSLS_resnet152_GeM_480_GCL.pth --backbone resnet152 --pool GeM --f_length 2048

This will produce the results on the MSLS validation set for this model. If you select --subset test, the file results/MSLS/test/MSLS_resnet152_GeM_480_GCL_predictions.txt will be generated. To evaluate the predictions you will need to submit this file to the MSLS evaluation server.

To apply PCA whitening run the apply_pca.py script with the appropriate parameters. For instance, for the example above, you have to run:

python3 apply_pca.py --dataset MSLS --root_dir PATH-TO-DATASET/MSLS/ --subset val --name MSLS_resnet152_GeM_480_GCL 

To reproduce our experiments we include a series of evaluation scripts in the 'scripts' folder, for the MSLS, Pittsburgh, Tokyo24/7, TokyoTM, RobotCar Seasons v2 and Extended CMU Seasons datasets. These scripts need the index files for each dataset that are available here and our model files, available here.

If you want to train a model on MSLS using the GCL function you must execute train.py with the appropriate parameters. For example:

python3 train.py --root_dir PATH-TO-DATASET/MSLS/ --cities val --backbone vgg16 --use_gpu --pool GeM --last_layer 2 

IMPORTANT NOTES FOR TRAINING

Make sure that the 'train_val' dir of the MSLS data set contains the graded files with the graded Ground Truth (the label files should be at the same level of the city folders in the train_val directory of MSLS).
Download the ground truth files from our DataVerse repository.
The password to extract the zip-file with the labels is 'gcl2022'.