Datasets

• The Cambridge dataset can be referenced at DFNet.
• The LIB and CON datasets were recorded using a smartphone. You can download the original videos of the LIB dataset at LIB, and the original videos of the CON dataset at CON.
• Dataset for Place recognition should put in UGNA-VPR/Cambridge/xxxx If you need datasets processed with Nerfsutido for NeRF training or for place recognition tasks, feel free to contact me.

Training Date for UE at dataset for UE. It should be put in data/, such as UGNA-VPR/data/dataset/Cambridge.

Pre-trained model

• Our pre-trained NeRF-h model at nerfh, it should be put in logs/, such as UGNA-VPR/logs/nerfh/shop1_4
• Our pre-trained VPR model at VPR-model, it should be put in logs/, such as UGNA-VPR/logs/MixVPR/Cambridge/ckpt_best.pth.tar
• Our pre-trained UE model at UE-model, it should be put in logs/, such as UGNA-VPR/logs/Cam_MixVPR_NBP
• our results and its pre-trained model at ours, it should be put in logs/, such as as UGNA-VPR/logs/tri_train_tea_1210_115010

Train VPR models

In training mode

self.parser.add_argument('---split', type=str, default='val', help='Split to use', choices=['val', 'test'])

# train the teacher net
python main.py --phase=train_tea

Train UE models

python train_npn.py -M [name of model] --setup_cfg_path config/Cambridge_training_setup.yaml

Test results

In test mode

self.parser.add_argument('---split', type=str, default='val', help='Split to use', choices=['val', 'test'])

needs to be changed to

self.parser.add_argument('---split', type=str, default='test', help='Split to use', choices=['val', 'test'])

python main.py --phase=test_tea	 --resume=logs/teacher_triplet/ckpt.pth.tar

The teacher_triplet/ckpt.pth.tar in the code needs to be changed to the appropriate name.