README.md

Notice

We recommend users download the image using the provided coordinates from the Google Earth API. Typically, each image is organized with the following naming format:

Country_Id_City_ULLon_ULLat_LRLon_LRLat.jpg

# Country: Name of the country
# Id: Unique identifier for this city
# City: Name of the city
# ULLon: Longitude of the upper left corner
# ULLat: Latitude of the upper left corner
# LRLon: Longitude of the lower right corner
# LRLat: Latitude of the lower right corner

⚠️ We recommand user to download from the data following data source:

• RSVQA: source
• NWPU: source
• RSICD: source
• RSITMD: source
• DIOR-RSVG: source
• RSVG: source
• UCM: source
• fMoW: source
• LLaVA: Follow instruction from here (we only use random 20K subset).

Pretraining

• Please download all the data through the Google Earth API and place it into a single directory with a name ending in “_Image.”

• Download our caption files from here and place all the json files in a folder named “OSMCapAnn.”

Finally, your pretraining data folder should be structured as follows:

|-PretrainData
|----XXX_Image
|    |---xxxxx.jpg
|    ...
|----OSMCapAnn
|    |features_01.json
|    ...

SFT

Stage2

• Download the instruction data from here and the corresponding images. Then organize the image folder names and json names in the following similar format:

|-Stage2Data
|----RSITMD_Image
|----RSITMD.json
|----RSITMDDetail_Image
|----RSITMDDetail.json
|----UCM_Image
|----UCM.json
|    ...

Stage3

• Download the instruction data from here and the corresponding images. Then organize the image folder names and json names in the following similar format:

|-Stage3Data
|----OSM_Image
|----OSM.json
|----LLaVA_Image
|----LLaVA.json
|    ...

Evaluation

• Classification, VQA are using the same format. Therefore, just download from source.
• Our reformat result of VG can be found at here.