Computer vision with AI

About this course

A course on computer vision as an engineering discipline: from pixels and image capture to a trained model that holds up beyond clean demo photos. It covers preprocessing, labeling, augmentation, classification and detection, per-condition metrics, and deploying a model to a device or server.

Five modules move from image-as-data through data collection and labeling, classification and detection, training and evaluation, to video and production. In "Data and labeling" you build a dataset of at least 300 images across four shooting conditions. In "Classification and detection" you implement one task both ways and compare. In "Training and evaluation" you run a full model audit with metrics broken out across at least three conditions.

For people willing to shoot and label their own images — there's no completing the projects without your own dataset. Requires Python and a basic grasp of machine learning. The course doesn't cover generative image models — the focus is classification, detection, segmentation, and video.

What you'll learn

• Choose preprocessing — scale, crop, normalization — to fit the task
• Build a dataset that prioritizes variety of shooting conditions over sheer volume
• Write labeling rules and work through edge cases
• Tell augmentations that help apart from ones that distort the data
• Solve one task as both classification and detection and compare results
• Read detection metrics — IoU, confidence threshold — and break down errors by condition
• Use transfer learning instead of training a model from scratch
• Design a production pipeline for video: frame rate, on-device vs. server, data drift

What you'll end up with

• Your own labeled dataset of at least 300 images across varied conditions
• One task implemented two ways — classification and detection
• A CV model audit with metrics broken out by shooting condition
• A production document for a video task: pipeline, frame rate, on-device vs. server (paid tiers include reviewer feedback and a certificate)

Not for you if you want image generation or a plug-and-play API without your own photos — here you shoot, label, and train the model yourself.