README.md

---
language:
- en
pretty_name: Geo Coordinate Augmented Google-Landmarks
task_categories:
- image-classification
source_datasets:
- Google Landmarks V2
size_categories:
- < 50K
license: cc-by-4.0
---


# Dataset Card for Geo Coordinate Augmented Google-Landmarks

Geo coordinates were added as data to a tar file's worth of images from the [Google Landmark V2](https://github.com/cvdfoundation/google-landmark). Not all of the 
images could be geo-tagged due to lack of coordinates on the image's wikimedia page. 

## Dataset Details

### Dataset Description

Geo coordinates were added as data to a tar file's worth of images from the [Google Landmark V2](https://github.com/cvdfoundation/google-landmark). There were many more images that could have
been downloaded but this dataset was found to be a good balance of data size and sample size. 

The intended use for the dataset was to demonstrate using a geo-filter in Qdrant along with a image similarity search. 

Not all of the images couuld be geo-tagged due to lack of coordinates on the image's wikimedia page.
We provide the raw geotagged file as a geojson document, train_attribution_geo.json. 
We also provide a json file that includes the data above along with embedding vectors for the images, id_payload_vector.json. 

Thingsvision was used as the library for creating the image embeddings with the following ThingVision model: 

```python
model_name = 'clip'
model_parameters = {
    'variant': 'ViT-B/32'
}
```

The code directory contains the Python code used to geotag the images as well as generated the vectors. It can also be used to upload the embeddings to a Qdrant DB instance.  This code is NOT for production and was 
more focused on quickly and correctly get the coordinates and embed the images. 

The license for this data and code match the license of the original Google Landmarks V2 Dataset: CC BY 4.0 license. 

## Uses

### Direct Use

The primary use is case is image similarity search with geographic filtering.