Skip to main content

Improving Your Model Over Time

Time: 5 minutes (reading)

Your inspection is deployed. Now how do you keep it performing well as conditions change, new defect types appear, or specs evolve?

The Library: Your continuous improvement tool​

Every capture the camera takes gets saved in the Library, along with the AI's prediction and confidence score. This is your gold mine for improvement.

Library view: browse captures, filter by prediction and confidence

Find what the AI got wrong​

  1. Go to the Library
  2. Browse recent captures
  3. Look for two things:
    • Misses: images where the AI's prediction is clearly wrong (checkbox in the top left of each thumbnail)
    • Low-confidence predictions: scroll down on any image to see confidence values. Low confidence = the AI was unsure
One mislabeled image can significantly hurt accuracy

In a small dataset, a single wrong label has an outsized impact. With only 5 training images, one mislabel corrupts 20% of your data. Always double-check every label before retraining, especially when your dataset is small.

Retrain with targeted data​

  1. Select the images the AI got wrong or was unsure about
  2. Click "Add to Active Recipe Train Set"
  3. Fix the labels if needed
  4. Click Retrain

Adding Library images to your training set

Focus on misses and low-confidence captures, not random new data. This is the most efficient way to improve.

Use Haystack to spot mislabels at scale

As your dataset grows, manually checking every label becomes impractical. Haystack lets you visually explore your training data, cluster similar images together, and quickly spot labels that look out of place. Run it periodically to keep your dataset clean.

Classifier improvement workflow​

  1. Review Library images → find errors and low-confidence predictions
  2. Add them to the training set
  3. Relabel if needed
  4. Retrain
  5. Use Haystack to explore your data visually and find mislabeled images at scale

Segmenter improvement workflow​

Segmenters take longer to label (pixel-level annotation), so there's a shortcut:

  1. Import problem images into the segmentation recipe
  2. Click Generate Predictions so the model pre-labels the new images as best it can
  3. Fix the predictions rather than labeling from scratch (much faster)
  4. Retrain

The philosophy​

  • Never stop iterating. The AI has high learning capacity, and it keeps improving with 50, 100, even 500+ images
  • Target failures specifically. Don't randomly add data. Add the cases where the AI struggles
  • Check for mislabels regularly. As your dataset grows, mislabels become harder to spot but still damage accuracy

Accelerate with GenAI tools​

Three AI-powered tools at tools.overview.ai can dramatically speed up your improvement cycles:

  • Defect Studio -- Generate photorealistic synthetic defect images up to 10,000x faster than waiting for real defects
  • Integration Builder -- Build production-ready Node-RED flows from plain English descriptions
  • AI Expert Helper -- Get 24/7 expert guidance on any camera question
The GenAI toolkit advantage

Together, these three tools can reduce your deployment time from days to hours. Generate synthetic training data instead of waiting for real defects, build integrations by describing what you want, and get instant expert guidance without waiting for support tickets.

→ See the full GenAI Toolkit guide