Definition
Team Swachh's model will learn from photos of bins. How many photos? Of which bins, in what light, in which seasons? Who takes them, and who decides which ones count as "overflowing"?
Planning the data means deciding, before collecting anything, exactly what data the solution needs, where it will come from, how it will be labelled, how much is needed, how it will represent every situation and group, and how it will be stored and protected. The result is a data plan.
In simple words
Before a cricket coach picks practice drills, they decide what skills the team needs and plan enough practice for each: batting, bowling, fielding, in different conditions. A data plan does the same for an AI model's practice examples.
Why it matters
The model can only be as good as its data. A careful plan prevents the most common capstone problems: too few examples, examples that all look alike, unclear labels and private information collected by accident.
Key terms
- Data plan
- A written plan for what data to collect, how, from where and how to protect it.
- Label
- The correct answer attached to an example, such as "overflowing" or "fine".
- Labelling guide
- Clear rules for deciding which label each example gets, so different people label the same way.
- Coverage
- How well the data includes every situation the model will meet: different bins, lighting, weather and angles.
- Balance
- Having a similar number of examples for each label.
- Data protection
- Keeping data secure and private: limited access, no unnecessary personal details, deleting when done.
Explained in detail
Start with a system map
Team Swachh first drew a system map. Bin fullness was affected by market days (+), festivals (+), the number of shops nearby (+), stray animals scattering garbage (+) and how recently the truck came (−). This told them that while their prototype would classify photos, a later version might also use market-day and festival calendars to predict which bins would fill next. For the capstone they kept the scope small: classify photos as overflowing or fine.
Team Swachh's data plan
| Question | Their answer |
|---|---|
| What data? | Close-up photos of public bins |
| Labels? | "Overflowing" (garbage above the rim or on the ground) and "fine" (below the rim) |
| How many? | At least 60 per label to start, aiming for 150 |
| Coverage? | All six bins, morning and evening, sunny and rainy days, different angles and phones |
| Who collects? | The team, plus two shopkeepers who volunteered |
| Privacy? | No people, faces or house numbers; delete any that slip through |
| Storage? | A shared folder only the team and teacher can open; deleted after the project |
| Test set? | One photo in five set aside and never used for training |
A labelling guide prevents disagreement
When Rohan and Tenzin labelled the same 20 photos separately, they disagreed on 5. Was a bin with garbage exactly at the rim overflowing or fine? Was a bin with one bag beside it overflowing? They wrote a labelling guide with clear rules and example photos for borderline cases: garbage above the rim or anything on the ground counts as overflowing. After that, they disagreed on only 1 photo in 20. Consistent labels matter, because a model learns whatever the labels teach it, including inconsistency.
Plan for exploration
- Count photos per label, per bin and per lighting condition, and fill gaps.
- Check for blurry, duplicate or mislabelled photos.
- Look for hidden shortcuts: if all overflowing photos are from one bin, the model may learn that bin instead of fullness.
Team Swachh also asked the two volunteer shopkeepers for consent in writing, explained exactly how the photos would be used, and promised to delete them after the project. They kept a simple log of who took each photo, when and of which bin, without any personal details, so that problems could be traced later. Planning these small details at the start saved them from collecting hundreds of photos that might later have had to be thrown away.
How it works
- Draw a system map and decide the exact prediction.
- Define the labels and write a labelling guide with examples.
- Decide how much data, and what coverage, you need.
- Plan collection, consent, privacy and storage.
- Set aside a test set and plan how you will explore and clean the data.
Try it yourself
The bin that taught the wrong lesson
Team Swachh collects their first 80 photos.
Good data plan choice or problem?
Sort each decision.
200 overflowing photos and 20 fine ones
A written labelling guide with borderline examples
Setting aside one photo in five for testing
Photos taken only on sunny afternoons
Saving photos that show passers-by's faces
Common mistakes
- Collecting data before deciding labels and coverage.
- Letting each person label by their own judgement, with no guide.
- Not checking for shortcuts such as one group appearing under only one label.
Check yourself
Summary
- A data plan decides what to collect, labels, quantity, coverage, privacy and storage.
- A labelling guide keeps labels consistent; a test set keeps evaluation fair.
- Count examples by group to catch gaps and hidden shortcuts.
Finished reading?
Sources: Digital Personal Data Protection Act, 2023 (MeitY) · CBSE Artificial Intelligence curriculum resources