Vision & Industrial Perception

Why this is hard

Academic vision assumes a balanced dataset. A production line gives you ten thousand good parts and eleven bad ones, under lighting that changes through the day, at a rate that will not wait for a round trip to a data centre. The research problem is learning from scarcity and staying calibrated once the process drifts.

Questions we are working on

1. How do you train a usable defect detector when the defect class has tens of examples, not thousands?
2. What does drift look like in a vision model before accuracy visibly falls, and can it be caught from the input distribution alone?
3. How much accuracy is lost bringing a model down to an edge accelerator at line rate, and where is that trade worth making?
4. Which frames are worth a human review, and how should uncertainty decide that rather than a fixed sample rate?

What comes out of it

• Anomaly and few-shot approaches for defect sets that are small and badly imbalanced
• Lighting-invariant preprocessing, camera calibration and multi-view geometry for real installations
• Quantisation and edge deployment measured against the accuracy actually lost, not the theoretical figure
• Active-learning loops that route uncertain frames to a person before they become a false accept

Where it ships