BSTLD stands for Bosch Small Traffic Lights Dataset. It is a vision-based traffic light detection dataset, which is of great significance for achieving fully autonomous driving in urban environments (identifying and detecting small targets in a large environment).

The dataset contains 13,427 images (resolution 1280×720) and 24,000 labeled traffic lights. The training set and test set contents are as follows:

Training set:

• 5,093 images
• Approximately once every 2 seconds
• 10,756 marked traffic lights
• Median traffic lights width: ~8.6 pixels
• 15 different labels
• 170 lights were partially blocked

Test set:

• 8,334 consecutive images
• Annotation at approximately 15 fps
• 13,486 annotated traffic lights
• Median traffic lights width: 8.5 pixels
• 4 labels (red, yellow, green, off)
• 2,088 lights were partially blocked

The dataset is provided by Heidelberg Image Processing Laboratory (HCL) and published at the 2017 International Conference on Robotics and Automation (ICRA).

Example data: