yanghongji2007/cross_view_localization_L2LTR

cross_view_localization_L2LTR

will be updated soon

Experiment Dataset

• CVUSA：https://github.com/viibridges/crossnet
• CVACT：https://github.com/Liumouliu/OriCNN

Dataset Preparation

Our method follow SAFA and DSM to get polar transform aerial images. Please download data_prearation.py for pre-processing.

Models

Pretrained model Download(Google's Official Checkpoint)

• R50+ViT-B_16 pretrained model

Our trained models for CVUSA and CVACT will be available soon

• Trained models

Train and test model

Train

python train.py --name CVUSA --dataset CVUSA --pretrained_dir YOUR_PRETRAINED_MODEL --output_dir MODEL_WHERE_WILL_BE_SAVE --dataset_dir YOUR_DATASET_PATH --learning_rate 1e-4 --weight_decay 0.03

Test

python test.py --name CVUSA --dataset CVUSA --output_dir MODEL_WHERE_WILL_BE_SAVE --dataset_dir YOUR_DATASET_PATH

If you want to use auto mixed precision, you should install APEX and add --fp16 in startup code.

Result

dataset	top-1	top-5	top-10	top-1%
CVUSA	94.05%	98.27%	98.99%	99.67%
CVACT_val	84.89%	94.59%	95.96%	98.37%