Awesome

Mapillary_Annotation

Using the API of the crowdsourcing platform Mapillary, we automatically download all available street-level images over the area of interest in the Netherlands for the year of 2017. Then, each downloaded image is matched with the corresponding LPIS object(s) it illustrates. We annotate images that are taken either towards the windshield direction (Case 1) or the window direction (Case 2).

We move the initial geo-location coordinates (lat1, lon1) to new coordinates (lat2, lon2) that are d = 10m away in the direction of angle θ.

For Case 1, we set θ = compass angle + 45 for the right half of the image and θ = compass angle−45 for left half.

For Case 2 we set θ = compass angle.

Consequntly, we use a No Reference Image Quality Assessment (NR-IQA) algorithm, namely BRISQUE, to remove bad quality images.

The code for downloading and annotating the images from the Mapillary platform is available in this link

The dataset is available in this link. It contains:

• street level images from the Mapillary API. We have also used the BRISQUE algorithm for No Reference Image Qaulity Assessment (NR-IQA) of the images, in order to exclude the unsuitable ones.

Example of good quality image

• a csv file, which contains the polygons of 45581 parcels together with an id for each parcel, the corresponging image id and date of the capture, the direction of the image that capture each parcel as well as the label of the illustrated crop. Below is shown an example of this file.

Distribution of Labels

Reference

If you use this dataset please cite the publication below

@inproceedings{sitokonstantinou2022datacap,
  title={DataCAP: A Satellite Datacube and Crowdsourced Street-Level Images for the Monitoring of the Common Agricultural Policy},
  author={Sitokonstantinou, Vasileios and Koukos, Alkiviadis and Drivas, Thanassis and Kontoes, Charalampos and Karathanassi, Vassilia},
  booktitle={International Conference on Multimedia Modeling},
  pages={473--478},
  year={2022},
  organization={Springer}
}