HighBuild-1M

HighBuild-1M is a multi-continental high-resolution benchmark dataset for single-view building height estimation from overhead imagery. Each sample contains a 1024 x 1024 RGB overhead image, a spatially aligned float32 building-height mask, and COCO-style building instance annotations.

1. Dataset Overview

HighBuild-1M covers six continents, 12 countries or regions, and 26 city groups. It contains 70,266 high-resolution RGB image tiles and 7,351,593 unique buildings, of which 6,616,041 have valid height attributes, representing 89.995% of all buildings.

Statistic	HighBuild-1M
Geographic coverage	6 continents
Countries or regions	12
City groups	26
RGB image tiles	70,266
Image dimensions	1024 × 1024 pixels
Unique buildings	7,351,593
Buildings with valid heights	6,616,041 (89.995%)
Building height annotations	32-bit floating-point TIFF, in meters
Building instance annotations	COCO JSON and instance-level source metadata JSON

Building counts are based on unique instances after cross-tile deduplication. A building may appear in multiple adjacent image tiles; therefore, the total number of per-tile instance records differs from the global count of unique buildings.

2. Dataset Contents

HighBuild-1M provides corresponding building height information, building footprint masks, and building instance annotations for each RGB image tile.

Each tile includes the following six data components:

Component	File Format	Description
RGB imagery	JPG	High-resolution optical overhead imagery
Building height map	Float32 TIFF	Pixel-level relative building heights in meters
Building footprint mask	UInt8 TIFF	Binary mask indicating building footprints covered by the source data
Height validity mask	UInt8 TIFF	Binary mask distinguishing pixels with valid building heights from other pixels
COCO instance annotations	JSON	Building instance polygons, bounding boxes, and associated attributes
Building instance source metadata	JSON	Building heights, validity information, original source identifiers, and instance geometries

2.1 Building Height Annotations

Building heights are represented as relative heights, consistently measured in meters.

Pixel values in the height TIFF files are defined as follows:

• Greater than 0: Building pixels with valid height attributes; pixel values represent building heights in meters.
• Equal to 0: Background regions.
• Equal to -1: Regions covered by building footprints but lacking valid height information.

The building footprint mask indicates the coverage of building footprints in the source data and should not be interpreted as a complete classification of all buildings and ground surfaces.

2.2 Building Instance Annotations

The dataset provides both COCO-format instance annotations and instance-level source metadata.

COCO annotations are provided for compatibility with commonly used instance segmentation tools and related research tasks. They include building footprint polygons, bounding boxes, areas, and selected building height attributes.

Instance-level source metadata additionally preserves building heights, source fields, original building identifiers, and within-tile geometry information. These records support source traceability, cross-tile instance analysis, and building height statistics.

The two annotation representations are generated through different processing pipelines. Instance counts and geometric representations may differ for certain tiles; therefore, they should not be treated as identical annotation records.

3. Data Organization and Splits

3.1 Directory Structure

The dataset is organized into training, validation, and test subsets. Each subset contains RGB imagery, building height maps and masks, and COCO instance annotations.

The directory structure is as follows:

HighBuild-1M/
└── data/
    ├── train/
    │   ├── images/
    │   ├── masks/
    │   └── coco_json/
    ├── validation/
    │   ├── images/
    │   ├── masks/
    │   └── coco_json/
    └── test/
        ├── images/
        ├── masks/
        └── coco_json/

Each directory preserves the original city identifiers and tile filenames, allowing RGB images, height rasters, and building instance annotations to be matched using the city name and tile ID.

3.2 Dataset Splits

A fixed training, validation, and test partition is provided for all 70,266 image tiles across the 26 city groups.

Subset	Number of Tiles	Proportion
Train	44,971	64%
Validation	11,242	16%
Test	14,053	20%
Total	70,266	100%

The partition was generated using a fixed random seed. The three subsets have mutually exclusive tile IDs and collectively cover the entire dataset.

4. Data Sources and Licensing

HighBuild-1M integrates high-resolution imagery, building footprints, and building height information from multiple publicly accessible geospatial data providers.

Because the source datasets are subject to different licensing and redistribution conditions, HighBuild-1M follows a source-specific licensing approach rather than applying a single unified license to all data components.

For source-specific license terms, required attribution, usage restrictions, and redistribution conditions, please refer to the HighBuild-1M Licensing Information (LICENSE.md)

Users are responsible for complying with the applicable license terms and attribution requirements of the corresponding data providers.