Erpix3lt/traffic-sign-classifier

Traffic Sign Classifier (GTSRB)

Model Summary

This is a traffic sign classification model trained on GTSRB classes. Architecture: lightweight CNN with feature extractor, 256-d bottleneck, and linear classifier.

This checkpoint is intended for research and analysis workflows, not safety-critical deployment.

Dataset

• Source: https://platform.ultralytics.com/maaaaaaaaaaaaaaaax/datasets/gtsrb-full
• Domain: German traffic sign recognition
• Image preprocessing:
  • Resize to 48 x 48
  • RGB input
• Training augmentation:
  • Random rotation: ±10 degrees
  • Color jitter: brightness 0.2, contrast 0.2

Model Architecture

• Backbone: custom CNN
• Block 1:
  • Conv(3,32), BN, ELU
  • Conv(32,32), BN, ELU
  • MaxPool2d(2), Dropout2d(0.2)
• Block 2:
  • Conv(32,64), BN, ELU
  • Conv(64,64), BN, ELU
  • MaxPool2d(2), Dropout2d(0.3)
• Block 3:
  • Conv(64,128), BN, ELU
  • Conv(128,128), BN, ELU
  • MaxPool2d(2), Dropout2d(0.4)
• Bottleneck:
  • Flatten
  • Linear(128 x 6 x 6 -> 256), ELU, Dropout(0.5)
• Head:
  • Linear(256 -> 42)

Framework and Weights

• Framework: PyTorch
• Weight format: state_dict checkpoint

Intended Uses

• Research on traffic sign recognition
• Transfer learning experiments
• Educational use for compact CNN pipelines

Out-of-Scope Uses

• Real-world safety-critical decision making

Limitations

• Trained on a narrow visual domain (GTSRB)