🎯 About this course
Computer vision succeeds or fails on data, evaluation and deployment. This course covers the working toolkit: image pipelines and label-preserving augmentation, convolutional backbones and transfer learning, object detection with IoU and NMS, segmentation with Dice and mIoU, precision–recall and mAP, and optimisation for cloud or edge.
Every topic ends in an auto-graded lab that runs on a laptop CPU with tiny synthetic images, and the applied studio has you fine-tune, evaluate and package a model with a model card.
👤Who it is for
Engineers and analysts who know basic deep learning and want to ship vision systems.
✨What makes it different
The emphasis is on evaluation honesty — source-based splits, per-class metrics, error analysis and on-device latency — not just training recipes.
By the end you will be able to
- Build image data pipelines with correct normalisation, label-preserving augmentation and leak-free splits.
- Explain CNN backbones, compute output shapes and parameter counts, and apply transfer learning with correct freezing and learning rates.
- Implement IoU and NMS and explain detector families and their trade-offs.
- Apply segmentation models and evaluate them with overlap metrics and appropriate losses.
- Evaluate vision systems with precision–recall, AP/mAP and structured error analysis, and choose operating thresholds.
- Optimise and deploy vision models within a latency budget on cloud or edge targets with human-review paths.
- Deliver a defensible vision model with a source-based split, error analysis, latency measurements and a model card.