Ranktuning

This is the official repository for the paper "Ranktuning: Improving Visual Plcae Recognition Preformance with Little Cost".

Getting Started

This repo follows the framework of Visual Geo-localization Benchmark for training and evaluation. You can refer to VPR-datasets-downloader to prepare test datasets.

The test dataset should be organized in a directory tree as such:

├── datasets_vg
    └── datasets
        └── pitts30k
            └── images
                ├── train
                │   ├── database
                │   └── queries
                ├── val
                │   ├── database
                │   └── queries
                └── test
                    ├── database
                    └── queries

Before training, you should download the pre-trained foundation model DINOv2(ViT-B/14) HERE and SALAD HERE.

Train

To train the model on MSLS

python3 train.py --eval_datasets_folder=/path/to/your/datasets_vg/datasets --eval_dataset_name=MSLS --foundation_model_path=/path/to/pre-trained/dinov2_vitb14_pretrain.pth --resume=/path/to/pre-trained/SALAD.pth --negs_num_per_query=4 --pos_num_per_query=4

Further finetuning on Pitts30k

python3 train.py --eval_datasets_folder=/path/to/your/datasets_vg/datasets --eval_dataset_name=Pitts30k --foundation_model_path=/path/to/pre-trained/dinov2_vitb14_pretrain.pth --resume=/path/to/resume/RankTuning_msls.pth --negs_num_per_query=2 --pos_num_per_query=2

Test

To evaluate the trained model on Pitts30k/MSLS:

python3 eval.py --eval_datasets_folder=/path/to/your/datasets_vg/datasets --eval_dataset_name=Pitts30k --resume=/path/to/resume/RankTuning_MSLS.pth or RankTuning_Pitts30k.pth

SALAD

Visual Geo-localization Benchmark

DINOv2