EOBuildings

This dataset is an ongoing effort to reproduce, improve and extends the INRIA Image Labelling dataset which is dedicated to buildings detection/segmentation from aerial imageries.

The original INRIA dataset reference are :

download url : https://project.inria.fr/aerialimagelabeling/files/

paper : Emmanuel Maggiori, Yuliya Tarabalka, Guillaume Charpiat and Pierre Alliez.“Can Semantic Labeling Methods Generalize to Any City? The Inria Aerial Image Labeling Benchmark” IEEE International Geoscience and Remote Sensing Symposium (IGARSS). 2017

The buildings footprints of the original INRIA dataset has only been made available for the five training AOI (Area of Interest) and on raster format. Somes issues has been noted by differents researchers on this dataset :

The annoation quality seems to be one of the main issues to obtains better models and compared them with enought confidence. Specifically in comparison with the AIRS dataset.

The raster building footprints formats is not compatible with instance segmentation task.

The more recents papers don't use the original test set, but instead use the original validation set to reports their performance metrics. So one of the strongest point of the dataset, to assess a model genericity (performance on unseen AOI during training), has been lost.

This version of the dataset, V0.1 contained :

All buildings footprints of the ten INRIA Area of Interset in in vector format (shapefile).

Two buildings footprints on 28km2 of a French department (Jura). One building footprints come from national buildings open data (French Mapping Institut) and the other is a mnaual annotation of the roof a seen on aerial ortho imagery.