Fine-Tuning Faster R-CNN for Guardrail Damage Detection

[View Project]

• Researched and implemented advanced fine‑tuning strategies on a Faster R‑CNN to detect guardrail damage, improving detection accuracy by 5% and decreasing false positives by 92% for Blyncsy, Inc.
• Designed a saliency‑scoring pipeline leveraging pre‑trained networks to automatically verify ground truth labels in the training and testing sets, automatically finding errors in roughly 10% of images.
• Engineered end‑to‑end training pipelines for a custom Faster R‑CNN, packaged as the BlyncsySFT pip package for easy deployment and reproducibility.
• Keywords: Python, PyTorch, Object Detection, pip