Evaluating Deep vs. Lightweight CNNs for Traffic Sign Recognition

Context

Traffic Sign Recognition for ADAS has to be both highly accurate and computationally efficient — a 99% accurate model is useless if it can’t run in real time on car-grade hardware. The paper quantifies the actual trade-off between a deep model (ResNet-50) and a lightweight one (MobileNetV3-Large) on the same dataset and same evaluation harness, instead of just citing headline accuracy numbers.

What I built

• Owned the ResNet-50 branch end-to-end as one of two student co-authors (alongside Yoel Augustan, supervised by Ivan Sebastian Edbert & Alvina Aulia).
• Implemented a two-phase transfer learning technique: train a new classification head on top of frozen base layers, then unfreeze and fine-tune on GTSRB.
• Prepared and augmented GTSRB so both models trained on identical data.
• Evaluated on accuracy, weighted F1, inference time on both CPU and GPU, and model complexity — not just headline accuracy.

Stack

Python · TensorFlow · Pandas · ResNet-50 · MobileNetV3-Large · GTSRB

Outcome

ResNet-50 reached 94.16% test accuracy with a 94.29% weighted F1; MobileNetV3-Large landed at 92.19% accuracy but was substantially faster on GPU, making it the better fit for real-time deployment. The paper was accepted and presented at ISCT — it’s published in the proceedings.