Sample Benchmark Datasets

id	id_train	ds_size	Urban Region	Country	Continent	Urban Type	Density	Roof Type	Download
1	364	399	Kakuma	Kenya	Africa	Refugee Camp	Sparse	Metal	Download
2	370	168	Denver	USA	America North	Peri-Urban	Grid	Shingles	Download
3	372	420	Montevideo	Uruguay	America South	Urban	Grid	Cement	Download
4	373	399	Montevideo Dense	Uruguay	America South	Urban	Dense	Cement	Download
5	391	231	Kutupalong	Bangladesh	Asia	Refugee Camp	Dense	Mixed	Download
6	394	504	Gornja Rijeka	Croatia	Europe	Rural	Sparse	Shingles	Download
7	397	756	Melbourne	Australia	Oceania	Urban	Grid	Cement	Download
8	398	152	Pemba	Tanzania	Africa	Rural	Sparse	Metal	Download
9	399	294	Christchurch	New Zealand	Oceania	Peri-Urban	Sparse	Shingles	Download
10	456	147	Pergamino	Argentina	America South	Peri-Urban	Grid	Mixed	Download
11	459	105	Silvania	Brazil	America South	Rural	Sparse	Shingles	Download
12	462	147	Desa Kulaba	Indonesia	Asia	Rural	Sparse	Metal	Download
13	463	168	Roseau	Dominica	America Central	Peri-Urban	Sparse	Mixed	Download
14	485	147	Pallabi Dhaka	Bangladesh	Asia	Urban	Dense	Mixed	Download
15	488	168	Dhaka	Bangladesh	Asia	Urban	Dense	Mixed	Download
16	489	147	Ggaba	Uganda	Africa	Peri-Urban	Dense	Mixed	Download
17	529	546	Inagi	Japan	Asia	Peri-Urban	Sparse	Mixed	Download
18	508	226	Tchiniambi	DRC	Africa	Peri-Urban	Dense	Metal	Download
19	539	672	Staraya Russa	Russia	Europe	Rural	Sparse	Mixed	Download
20	530	189	Banyuwangi	Indonesia	Asia	Urban	Dense	Shingles	Download
21	526	672	Dzaleka	Malawi	Africa	Refugee Camp	Dense	Metal	Download
22	523	252	Bogota	Colombia	America South	Urban	Grid	Mixed	Download
23	524	315	Soudure	Niger	Africa	Rural	Dense	Mixed	Download
24	525	420	Quincy	USA	America North	Peri-Urban	Grid	Shingles	Download
25	528	756	Ngaoundere	Cameroon	Africa	Peri-Urban	Grid	Metal	Download

Source and Credit : Anna Zanchetta , More details

How to Get the Data

You can use the following Python script to generate download links for all available training datasets:

# List of training IDs from the table
training_ids = [
    364, 370, 372, 373, 391, 394, 397, 398, 399, 456, 
    459, 462, 463, 485, 488, 489, 529, 508, 539, 530, 
    526, 523, 524, 525, 528
]

base_url = "https://fair-dev.hotosm.org/api/v1/workspace/download/training_{}/preprocessed/preprocessed.tar.xz/"

download_links = [base_url.format(train_id) for train_id in training_ids]

for link in download_links:
    print(link)

Training details can be found from API call like this : https://fair-dev.hotosm.org/api/v1/training/364/

Dataset Structure & Download Guide

Overview

This dataset consists of 256x256 pixel image tiles that follow the Mercator tiling scheme. Each tile is associated with:

• Imagery ("chips/")
• Vector labels ("labels/")
• Binary masks ("binarymasks/") (optional)

The filenames follow this naming convention:

OAM-{mercantile_tile_x}-{mercantile_tile_y}-{zoom_level}.ext

Where:

• OAM: Prefix indicating OpenAerialMap (or similar sources).
• mercantile_tile_x / mercantile_tile_y: Tile coordinates in the Mercator grid.
• zoom_level: The zoom level of the tile.
• ext: .tif (imagery), .geojson (labels), .mask.tif (binary masks).

---

Folder Structure

1. Chips (chips/)

• Contains satellite/aerial imagery tiles.
• Each .tif file corresponds to a specific Mercator grid tile.
• Example: OAM-1251460-1026614-21.tif.

2. Labels (labels/)

• Contains vector annotations (GeoJSON).
• Each file is clipped to the exact boundary of the corresponding image tile.
• Example: OAM-1251450-1026604-21.geojson matches OAM-1251450-1026604-21.tif.

3. Binary Masks (binarymasks/)

• Rasterized (burned) version of the labels.
• Binary format (0/1) indicating building footprints.
• Not required if users prefer to generate masks from chips/ and labels/.
• Example: OAM-1251456-1026606-21.mask.tif.

Distribution of datasets

More info

Check this presentation for Anna's research result and explanation here