GPG2A (Geometric Preserving Ground-to-Aerial image synthesis)

Official code repository for the paper "Cross-View Meets Diffusion: Aerial Image Synthesis with Geometry and Text Guidance" presented at WACV, 2025.

Prerequisites

Training & Testing

• accelerate
• diffusers
• Pytorch >= 2.1.2
• torchvision >= 0.16.2
• pillow
• tqdm

Evaluation (Additionaly)

• scikit-image
• opencv
• lpips

Dataset

Please fill out the form provided here to download VIGORv2

Pre-trained Weights

Our checkpoints can be downloaded from the following link

Training

After downloading VIGORv2 and the pre-trained weights, for the best experience, make sure your repository tree is as follows:

GPG2A
├── checkpoints
│   ├── Stage_I/
│   └── Stage_II/
├── data
│   └── VIGORv2/
├── Stage_I/
├── Stage_II/
└── utils.py

Stage I (BEV Layout Estimation)

To train stage I (BEV Layout Estimation) navigate to the root directory and run the following script

bash ./Stage_I/train.sh

Note that you can customize the training experiment using the arguments provided in ./Stage_I/train.py by simply passing them to ./Stage_I/train.sh

This script will save checkpoints at every epoch in ./OUTPUT_DIR/checkpoints

Stage II (Diffusion Aerial Synthesis)

To train stage II (Diffusion Aerial Synthesis) navigate to the root directory and run the following script

bash ./Stage_II/train.sh

Note that you need to have the checkpoints of stage I saved in the correct location (as shown in the above tree).

You can customize the training experiment using the arguments provided in ./Stage_II/train.py by simply passing them to ./Stage_II/train.sh, for example:

--text_type="dynamic"
--VIGOR_mode="samearea"

The following are the main experiments ran in the paper:

Argument	Experiment
--text_type="dynamic"	Dynamic text prompt extracted from Gemini discriptions of the ground image
--text_type="raw"	Gemini descriptions of the ground image without processing
--text_type="constant"	A constant single generic prompt, e.g., "High-quality aerial image"
--text_type="city"	Same as 'constant' prompt but with city name
--text_type="empty"	Empty string
--VIGOR_mode="samearea"	Same-area training protocol
--VIGOR_mode="crossarea"	Cross-area training protocol
--f_bev	Use the raw BEV feature instead of the estimated layout map

Testing

Stage I (BEV Layout Estimation)

To test stage I (BEV Layout Estimation) navigate to the root directory and run the following script

bash ./Stage_I/test.sh

Note that you need to have the checkpoints of stage I saved in the correct location (as shown in the above tree).

Stage II (Diffusion Aerial Synthesis)

To test stage II (Diffusion Aerial Synthesis) navigate to the root directory and run the following script

bash ./Stage_II/test.sh

Note that you need to have the checkpoints of stages I and II saved in the correct location (as shown in the above tree).

This script saves all generated images of the test set in the specified output directory (./log_imgs_test by default)

Evaluations

This section covers evaluating the generated images from GPG2A, so it assumes already having a directory of generated images (from the test set).

All evaluation scripts are in ./Evaluation/

Pre-trained SAFA model

Please download the pretrained SAFA model here and extract to the same directory.

Generate aerial images

Please follow the training and testing tutorial above to generate aerial images for evaluation with correct naming and directory.

$Sim_c$ and $Sim_s$

To calculate $Sim_c$ and $Sim_s$, please run the command below,

python eval_SAFA.py --model_path SAFA_PRETREAINED_MODEL_PATH --experiment_name YOUR_PREFERRED_EXPERIMENT_NAME --image_path GENERATED_IMAGE_PATH --gt_path GROUND_TRUTH_IMAGE_PATH

For example,

python eval_SAFA.py --model_path ../GPG2A_eval_SAFA_model/ --experiment_name GPG2A_experiment --image_path ../log_imgs_test/ --gt_path ../Data/VIGOR/

FID-SAFA

To calculate FID-SAFA, please run the command below,

python eval_FID.py --model_path SAFA_PRETREAINED_MODEL_PATH --experiment_name YOUR_PREFERRED_EXPERIMENT_NAME --image_path GENERATED_IMAGE_PATH --gt_path GROUND_TRUTH_IMAGE_PATH

For example,

python eval_FID.py --model_path ../GPG2A_eval_SAFA_model/ --experiment_name GPG2A_experiment --image_path ../log_imgs_test/ --gt_path ../Data/VIGOR/

LPIPS, SSIM, PSNR

To calculate LPIPS, SSIM, PSNR, please run the command below,

python test_lpips.py --experiment_name YOUR_PREFERRED_EXPERIMENT_NAME --image_path GENERATED_IMAGE_PATH --gt_path GROUND_TRUTH_IMAGE_PATH

For example,

python test_lpips.py --experiment_name GPG2A_experiment --image_path ../log_imgs_test/ --gt_path ../Data/VIGOR/