Deep Vision

• CNNs, residual networks and vision transformers compared
• Transfer learning, freezing schedules and learning rate selection
• Class imbalance, hard example mining and reading a confusion matrix

Tasks

• Object detection with YOLO-family and two-stage detectors, and mAP interpretation
• Semantic and instance segmentation and the IoU metrics that govern them
• Tracking, pose estimation and OCR as common production requirements

Engineering

• ONNX export, quantisation and inference on edge hardware
• Latency, throughput and batch size trade-offs under a real load
• Serving, versioning and rollback when a new model regresses

Certification

• An end-to-end project from data collection through deployment
• Monitoring for drift and building a retraining pipeline
• Documentation, model cards and defending design decisions under review