Skip to content

Prediction-Assisted Labeling

Prediction-assisted labeling accelerates your workflow by using model predictions to bootstrap new labels.

Benefits:

  • Faster labeling — Correcting a mostly-correct prediction is quicker than labeling from scratch
  • Targeted feedback — See where your model succeeds and fails, helping you choose the most useful frames to label

The Active Learning Loop

┌─────────────────────────────────────────────────────────┐
│                                                         │
│   Label frames  →  Train model  →  Predict  →  Correct  │
│        ↑                                          │     │
│        └──────────────────────────────────────────┘     │
│                                                         │
└─────────────────────────────────────────────────────────┘
  1. Label a small set of frames manually
  2. Train a model on your labels
  3. Predict on suggested or unlabeled frames
  4. Correct the predictions to create new training data
  5. Repeat until accuracy is satisfactory

Understanding the Seekbar

After running inference, the seekbar shows different markers:

Marker Meaning
Thick black line Manually labeled frame
Thin black line Frame with predictions
Dark blue line Suggested frame with manual labels
Light blue line Suggested frame with predictions
Red marker Suggested frame ready for review

Correcting Predictions

Predicted instances appear in grey with yellow nodes. To edit them:

  1. Double-click the predicted instance to convert it to an editable instance
  2. Adjust the nodes as needed
  3. Save your changes

Fixing predictions

Visual feedback

After converting a prediction:

  • Red nodes = unchanged from prediction
  • Green nodes = manually adjusted

This helps you track which nodes you've reviewed.

Predictions don't train automatically

Predicted instances are not used for training until you convert and correct them. Always double-click to convert before making edits.

Best Practices

  1. Generate new suggestions regularly — Active learning works best with fresh suggestions based on your latest model

  2. Focus on failure cases — Prioritize frames where predictions are wrong or uncertain

  3. Don't over-correct — If a prediction is close enough, a small adjustment is fine

  4. Iterate frequently — Several rounds of train → predict → correct typically yields better results than one large labeling session

Next Steps

Once you have accurate frame-by-frame predictions, you're ready to:

  • Run inference on full videos — Predict poses across entire clips
  • Track identities — Link instances across frames (see Tracking methods)
  • Proofread tracks — Use the proofreading tools to fix tracking errors