Validate and Diagnose

Run validation

yolo val model=runs/detect/forklift_v1/weights/best.pt data=my_dataset/data.yaml

That re-runs the val set and writes new artifacts to runs/detect/val/. The most useful files:

Read the confusion matrix first

The confusion matrix is the single most useful diagnostic. It's a square plot, classes on both axes, with the diagonal showing correct predictions and the off-diagonal showing confusion.

              predicted
              forklift  person  pallet  background
ground   ┌─────────────────────────────────────────┐
forklift │   0.78      0.04    0.10      0.08
truth    │
person   │   0.03      0.91    0.00      0.06
pallet   │   0.18      0.00    0.65      0.17
back-    │   0.04      0.07    0.05      0.84
ground   └─────────────────────────────────────────┘

Reading this:

• forklift → pallet cell is 0.10 — 10% of forklift instances are predicted as pallet. Class confusion.
• pallet → background cell is 0.17 — 17% of pallets are missed entirely. Recall problem.
• background → forklift is 0.04 — small number of false positives where the model sees a forklift in empty space.

The fix list writes itself:

1. Add hard examples that disambiguate forklift from pallet (forklifts visible behind pallets, partial forklifts, pallets close to forklifts).
2. Add more pallet examples — 17% missed is too high.
3. Background hard negatives — empty corridors that look like forklift cluttered backgrounds.

Read the per-class AP

Per-class AP from the val log:

                 Class   Images   Instances     Box(P    R    mAP50    mAP50-95)
                   all      50         102      0.78  0.71    0.74        0.51
              forklift      30          45      0.85  0.78    0.81        0.58
                person      40          50      0.92  0.88    0.91        0.64
                pallet      20           7      0.58  0.48    0.51        0.31

pallet is way below the others — and it has only 7 instances in val. Either the model is bad at pallets or you don't have enough val data to tell. Both fixes lead to "collect more pallets."

Are you overfit?

Open results.png. The training and validation data loss curves should track each other roughly. If training loss keeps falling while validation loss starts climbing, you've overfit — the model is memorizing. The opposite failure — both curves stuck high — is underfitting, and the bias-variance tradeoff is the lens for both.

Mitigations:

• Reduce epochs (the patience arg auto-stops training when val mAP plateaus, which usually catches this).
• Add data augmentation — hsv_h, hsv_s, hsv_v, degrees, translate, scale, fliplr, mosaic. Ultralytics YOLO defaults are aggressive; tone them down only if you have a reason.
• More data — almost always the right answer.

Visual sanity check

Compare val_batch0_labels.jpg and val_batch0_pred.jpg side by side. The differences are your failures, in concrete pictures. Patterns in those pictures (consistent missing class, consistent wrong region) are dataset stories — and dataset stories beat hyperparameter tuning every time.