Basic Traffic Sign Detection Dataset

Dataset for training traffic sign classification and detection models.

This dataset was created to identify specific Brazilian regulatory traffic signs. Due to the universal nature of road signage, it may also be applicable to other contexts. Specifically, the dataset includes:

• R-26 (straight foward);
• R-25a (turn left);
• R-25b (turn right);
• R-3 (no straight ahead);
• R-4a (no left turn);
• R-4b (no right turn).

It was composed of collected and labeled images, added to others present and already labeled in four other datasets of the Roboflow Universe platform. The images of the classes of interest were separated, and data augmentation was performed using this first small quantity of images. In the beggining, augmentations were applied through the aforementioned platform, and then more through external and customized code.

Datasets Used

1. Dataset: placas-de-transito-sm6dy

Trabalho. (2025). placas-de-transito Dataset [Dataset]. Roboflow Universe. https://universe.roboflow.com/trabalho-c3kzw/placas-de-transito-sm6dy

2. Dataset: Deteco-placasV2

detectplacas. (2025). Deteco-placasV2 Dataset [Dataset]. Roboflow Universe. https://universe.roboflow.com/detectplacas/deteco-placasv2

3. Dataset: placas-pg0mx

ictarcisio. (2025). placas Dataset [Dataset]. Roboflow Universe. https://universe.roboflow.com/ictarcisio/placas-pg0mx

4. Dataset: tccCarroAutonomoV2

tcc. (2024). tccCarroAutonomoV2 Dataset [Dataset]. Roboflow Universe. https://roboflow.com

Roboflow Augmentations

• Horizontal and vertical shear between -15° and +15°
• Random brightness adjustment between -38% and +38%
• Random exposure adjustment between -16% and +16%
• Gaussian blur between 0px and 8px
• Salt and pepper noise in 10% of pixels

Code Augmentations

• Standardized size for 700 x 700 pixels and 640 x 640 cropping in 50% of samples
• Constant brightness adjustments up to 15%
• Frequent (50% chance) simple or Gaussian blur
• Small scaling changes
• Optional rotations between 15° and 30°
• Synchronization with YOLO labels, using minimum visibility filters of 10% and a sanitization function