Pedestrian-Traffic-Lights (PTL)

Dataset Overview

Dataset Name

ImVisible: Pedestrian Traffic Light Dataset

Dataset Content

• Image Data: Street‑intersection images annotated with pedestrian traffic light colors and crosswalk locations.
• Image Characteristics: Diversity in weather, location, orientation, and intersection size and type.

Statistics

• Total Images: 5,059
• Split:
  • Training: 3,456 (68.3 %)
  • Validation: 864 (17.1 %)
  • Test: 739 (14.6 %)

Label Information

• Format: For each image, labels include filename, class, coordinates (x1, y1, x2, y2) and occlusion flag.
• Class Definitions:
  • 0: Red
  • 1: Green
  • 2: Countdown Green
  • 3: Countdown Blank
  • 4: None
• Class Distribution:
  • Red: 1,477 (29.2 %)
  • Green: 1,303 (25.8 %)
  • Countdown Green: 963 (19.0 %)
  • Countdown Blank: 904 (17.9 %)
  • None: 412 (8.1 %)

Download

• Training Images: 876×657 resolution
• Validation & Test Images: 768×576 resolution
• Full‑Resolution Dataset: 4032×3024 resolution 1 and 4032×3024 resolution 2

Model Information

• Model Name: LytNet
• Function: Recognize traffic‑light colors and predict crosswalk positions.
• Performance:
  • LytNet V1: Accuracy – Red 0.97, Green 0.94, Countdown Green 0.99, Countdown Blank 0.86; Recall – Red 0.96, Green 0.94, Countdown Green 0.96, Countdown Blank 0.92; Angle error 6.27°; Start point error 0.0763; End point error 0.0510.
  • LytNet V2: Accuracy – Red 0.98, Green 0.95, Countdown Green 0.99, Countdown Blank 0.93; Recall – Red 0.96, Green 0.96, Countdown Green 0.97, Countdown Blank 0.97; Angle error 6.15°; Start point error 0.0759; End point error 0.0477.

Application

• App Type: iOS demo app
• Function: Runs the LytNet model to output traffic‑light colors and crosswalk locations.
• System Requirement: iOS 11 or later.