stop-sign-automl

AutoML Neural Network Model for Stop Sign Classification

Model Summary

This model was trained using AutoGluon MultiModalPredictor (v1.4.0) on the dataset ecopus/sign_identification. The task is binary image classification, predicting whether a stop sign is present (1) or absent (0) in the input image.

• Best Model: AutoML-selected neural architecture (Hybrid CNN/Transformer backbone via AutoMM)
• Validation Strategy: Stratified 80/20 train/test split with early stopping on validation
• Precision / Recall / F1: Reported in confusion matrix and classification report

---

Dataset

• Source: ecopus/sign_identification
• Size: ~X samples (replace with your count)
• Features:
  • image: stop sign or non-stop sign photo
  • label: binary class (0 = no stop sign, 1 = stop sign present)

---

Preprocessing

• Images saved as .png files from dataset byte arrays
• Train/test split stratified on label
• AutoGluon applies default image preprocessing:
  • Resizing to fixed resolution
  • Normalization
  • Default augmentations (random crop/flip/resize)

---

Results

Test Metrics (example, update with actual numbers)

• Accuracy: 0.94
• Precision: 0.93
• Recall: 0.94
• F1: 0.94

Confusion Matrix

Balanced classification with a small number of false positives/false negatives.

---

Error Analysis

• Misclassifications often occur with:
  • Occluded or partially visible stop signs
  • Unusual lighting conditions (night, glare)
  • Red objects mistaken for stop signs (background clutter)

---

Intended Use

• Educational use only
• Demonstration of AutoML for neural networks in CMU course 24-679
• Not suitable for deployment in safety-critical systems

---

Limitations

• Performance may degrade on images outside the dataset distribution
• Sensitive to dataset bias (lighting, camera angle, geography)
• May fail in adversarial conditions (graffiti, damaged signs)

---

License

• MIT

---

Hardware/Compute

• Training performed on Google Colab with a T4 GPU
• AutoML time budget: 30 minutes (1800s)

---

AI Usage Disclosure

• This model was built using AutoGluon AutoML framework
• Hyperparameter and architecture search were automated