Traffic Sign Detection (CV)

Existing detection systems struggled with class imbalance (163:1) and could not distinguish between visually similar signs

Built a two-stage cascade architecture: YOLO for coarse detection (5 groups) + 5 specialized ResNet-18 models for fine classification. Achieved 94.07% mAP50 and 91.71% recall — a 12.7% improvement over single-stage baseline. Real-time WebSocket inference with browser camera support

Achieved 94.07% mAP50 (+12.7% over baseline) Real-time inference via WebSocket with browser camera Reduced detection errors on confusing sign classe