Start With the Business Objective
Translate a business goal into a vision task, a class list, and a success metric before any image is collected.
A common reason datasets fail in production: the team started collecting images before agreeing on what the model was actually supposed to do. The cure is a written objective. Done in 30 minutes; saves months. Aligns with the steps of a CV project guide.
Write a one-page objective that names the business goal, the vision task, the classes, and the metric that defines success.
Business goal in one sentence.
Vision task: detect / segment / semantic / classify / pose / OBB.
Final class list (3–10 classes is a good first target).
Success metric: e.g. mAP@0.5 ≥ 0.80 on the holdout set, or recall ≥ 0.95 on safety class.
Hands-on
Link to this sectionThe four-line spec#
A vision project that ships answers four questions in writing — see defining project goals and the broader data collection and annotation guide:
| Field | Example |
|---|---|
| Business goal | "Reduce dock-door incidents by 30% in 6 months." |
| Vision task | Object detection — boxes for forklifts and people. |
| Classes | forklift, person, pallet (background = no label). |
| Success metric | mAP@0.5 ≥ 0.80 and recall on person ≥ 0.95 on a holdout from cameras the model has never seen. |
Pin these to a wiki. Every dataset and labeling decision below traces back to this page.
Link to this sectionPicking the task is picking the labels#
Each vision task implies a different annotation cost — see Ultralytics YOLO tasks for definitions:
| Task | Labels | Relative annotation cost |
|---|---|---|
| Classification | One label per image | 1× |
| Detection | Class + axis-aligned box per object | 3–5× |
| OBB | Class + rotated box | 4–6× |
| Pose | Class + keypoints per object | 5–8× |
| Instance segmentation | Class + pixel mask per object | 8–15× |
| Semantic segmentation | Class ID per pixel | 8–20× |
Pick the simplest task that solves the business problem. Most teams reach for segmentation when detection would do — and run out of annotation budget. The CV Foundations course covers the task-picking decision tree in detail.
Link to this sectionCoarse vs. fine class counts#
When you write the class list, decide between coarse and fine granularity:
- Coarse ("vehicle", "non-vehicle"): cheaper to label, faster to converge, less informative.
- Fine ("sedan", "SUV", "pickup", "motorcycle"): more useful downstream, more expensive, harder to keep consistent.
Start with what the business actually needs. You can always merge fine classes into coarser groups later (one line of code); you can't split coarse labels back into fine ones without re-labeling.
"Anything that looks dangerous" or "stuff blocking the aisle" is not a class — it's a request for a judgment call from the annotator. Models can't learn from inconsistent judgment. If the class is hard to write down, it'll be hard to label consistently.
Link to this sectionDefine the metric before you train#
Pick the metric that matches the business goal: a safety system cares most about recall on the safety class, a counting system about balanced per-class AP, and a sorting system about precision so it doesn't ship the wrong item. The YOLO performance metrics guide covers what each metric means.
A useful exercise: write the acceptance test in plain English. Example: "On 500 frames from cameras the model has never seen, recall on person must be ≥ 0.95 with precision ≥ 0.85." That sentence drives the rest of the course.
Link to this sectionSet expectations early#
Write down what the model is not expected to do, too: the look-alikes you'll tolerate it confusing, the conditions you won't promise (pitch dark, heavy occlusion), and the long-tail classes that are out of scope for v1. No model is 100% accurate, so naming the acceptable misses up front keeps a model that hits its target from being judged against an unwritten 100% bar.
Write your project's four-line spec on a single page: business goal, vision task, class list, success metric. Share it with the stakeholder who's paying for the model. If they push back on any line, that's the line to fix before doing anything else.
Your business goal fits in one sentence.
You've picked exactly one vision task.
Your class list is 3–10 specific, non-overlapping classes (or you've written down why you need more).
Your success metric is a number, on a holdout, with a precision/recall constraint.
We have a target. Next: write the dataset spec that says what images we need to hit it.