Dataset Quality Control

The seven-check QC pass

Run all seven before any training run. None of them take more than 10 minutes — the preprocessing annotated data guide covers many of these in detail:

Spot-check by sampling

You don't need to review every image. A 100-sample stratified audit catches systemic issues:

• 10 from each class, picked at random.
• 10 background images.
• 10 "edge cases" from your spec.
• The lowest 10 confidence predictions from a quick pretrained-model run on the train set.

The last bucket is gold: it surfaces the images the model finds unusual — usually mislabeled or genuinely odd.

Class balance: what to do when it's broken

Real datasets are imbalanced. The question is whether it'll hurt training.

   forklift  ████████████████████████ 8000 instances
   person    ████████████████ 5400
   pallet    ██ 700                              ← red flag

A 10:1 imbalance like the one above will produce a forklift-confident model that misses pallets. Three fixes, in order of preference:

1. Collect more pallets. Always the best answer. Aim for ≥ 1500 images and ≥ 10,000 labeled instances of every class, not just the easy ones.
2. Oversample at training time. Augment minority-class images more aggressively so each epoch sees them more often.
3. Class-weighted loss (advanced). Don't reach for this until 1 and 2 are exhausted.

Dedup once more, after labeling

Even after upload-time dedup, post-labeling dedup catches a different kind of duplicate: visually distinct frames with identical labels. They're usually:

• Two camera angles of the same event with the same labels.
• Sequential video frames where nothing moved.

Drop them — they pad volume without adding training signal.

The leakage check is not optional

Random per-image splits leak whenever images are correlated (same scene, same camera, same day). Lesson 6 covers the splitting strategy in depth — but the QC version of the question is:

Does the same scene / camera / day appear in both my train and val splits?

If yes, your val mAP is overstating reality, and the model that ships will be worse than the one you measured. The fix is a grouped split — train and val never share a scene-key. The preprocessing annotated data guide walks through the mechanics; the splits lesson next covers the strategy.

Why fixing labels beats tuning hyperparameters

A common pattern: model val mAP is 0.62, team spends two weeks on hyperparameter tuning, mAP gets to 0.64. Then someone audits the data and finds 8% of forklift labels are mis-classified as pallet. They fix the labels and retrain — mAP jumps to 0.78.

This is the rule of thumb: before changing anything else, fix the data. The fine-tuning guide, the YOLO performance metrics guide, and the preprocessing annotated data guide all reinforce it. Models are good at generalizing from clean signal; they can't generalize from noise.

Document the QC result

Before training, write a one-page QC report:

Dataset: forklift_v1
Audit date: 2026-05-09
Total images: 4,820   |   Total labels: 18,442
Background images: 380 (7.9%)
Per-class instances:
  forklift  6,801
  person    8,310
  pallet    3,331
Visual audit: 120 images sampled, 4 mislabeled (3.3%), fixed.
Duplicates removed: 312
Leakage check: split-by-day, no scene overlap.
Class imbalance: pallet ~2.5× lower than person; ok, not collecting more this round.
Ready to train: yes.

This report is also the artifact you hand to the auditor or the customer. "Yes, we audited the dataset" with no document is not the same as "yes, here's the audit."