System
Vision task
One production-shaped task: data, train or adapt, eval, a small API.
Computer Vision
Images as data: classification, detection, segmentation, and modern vision-language — with evals that match the job, not ImageNet folklore.
Created by Baljeet Dogra
Course Objectives
- Build a dataset and a training pipeline that does not leak.
- Choose classification, detection or segmentation because of the product.
- Evaluate with the metric the job uses, not top-1 by default.
- Know when a VLM is a tool and when it is theatre.
What you will produce
Eval
Metric card
Why this metric, error slices, and known failure modes.
6 weeks curriculum
Expand a part for the syllabus. Content stays searchable when closed.
01 Pixels to tensors Week 1
Colour, augmentations, leakage via resize, annotation quality.
- Pipelines
- Augment
- Labels
- Leakage
02 Classify and retrieve Week 2
Backbones, transfer, embeddings for image search.
- Backbones
- Transfer
- Embeddings
- Errors
03 Detect and segment Weeks 3–4
Boxes, masks, IoU, when a heatmap is not a box.
- Detection
- Segmentation
- IoU
- Post-process
04 VLM and ship Weeks 5–6
Vision-language models, grounding, a small service. Capstone.
- VLMs
- Grounding
- Serving
- Capstone
Who this is for
ML engineers
You have trained tabular or text models. Images have different failure modes.
Product engineers in vision-heavy domains
Inspection, documents, retail shelves — you need the real metrics.
Prerequisites
- Deep Learning Specialisation or equivalent
- Python
- GPU access for training weeks
Not a fit if
- Photoshop filters as “computer vision”
- People with no training-loop experience
Related: Deep Learning Specialisation · NLP Advanced Techniques
Questions
OpenCV or deep models?
Both. Classical ops still earn their keep in pipelines. The course is honest about when.
Will we label data?
Yes. Bad labels are the usual failure, not the backbone.
Ready to start?
Six weeks to a vision task with a metric that matches the job. Create an account to enrol.
Enrol now