Hong Kong Geospatial Deep Learning Datasets

Image Sources

• Aerial Photography: Images captured at flying height of about 12,000 feet
• Resolution: 0.25m per pixel

The orthophotos are provided by the Lands Department of the Government of the Hong Kong Special Administrative Region.

For the usage of aerial images, please follow the Terms and Conditions of CSDI Portal.

Annotation License

The annotations of bounding boxes and segmentation masks are under the MIT License.

Object Detection Categories (8 Classes)

The object detection dataset includes annotations for the following features:

Segmentation Categories (2 Classes)

The semantic segmentation dataset provides pixel-level annotations for:

YOLO Format for Object Detection

Format Structure:

• One text file per image
• Each line represents one bounding box
• Format: class_id x_center y_center width height
• All values are normalized (0-1) relative to image dimensions

Example:

# Example annotation for an image with 3 objects
# Format: class_id x_center y_center width height
2 0.342 0.551 0.123 0.225   # Container lot
7 0.658 0.237 0.042 0.087   # Vessel
6 0.518 0.482 0.025 0.035   # Vehicle
                    

YOLO Format for Segmentation

Format Structure:

• One text file per image
• Each line represents one polygon
• Format: class_id x1 y1 x2 y2 ... xn yn
• All coordinates are normalized (0-1)

Example:

# Example segmentation for a building
# Format: class_id x1 y1 x2 y2 ... xn yn
0 0.123 0.456 0.133 0.467 0.153 0.485 0.183 0.492 0.198 0.472 0.173 0.458 0.123 0.456

# Example segmentation for a road
1 0.301 0.617 0.312 0.645 0.354 0.683 0.385 0.674 0.367 0.629 0.328 0.603 0.301 0.617