Skip to main content

Image Recognition Accuracy

CoolR consistently achieves an image recognition accuracy rate of over 96%, and typically 98% or above when at least 4+ good quality images of each product are available.

Image Recognition Accuracy Methodology​

Our accuracy rates are calculated using a rigorous testing process conducted every two weeks. We randomly select 1,000 images from different customers and industries, containing approximately 40,000 total product facings. Each result undergoes manual verification to ensure precision.

Accuracy Calculation​

We use a strict scoring system where only correctly identified products count toward accuracy. For example, if we detect 10 products in an image:

  • 8 correctly identified = +8 points
  • 1 incorrectly identified (false positive) = -1 point
  • 1 missed product = -1 point

Result: 80% accuracy (8/10)

A 96% accuracy rate represents 38,400 correctly identified products out of 40,000 total, with both missed detections and misclassifications penalized equally.

Why This Method Ensures Reliability​

  • Quality Control: Images with poor lighting, washout, or other quality issues that would unfairly skew results are excluded from testing.
  • Cross-Industry Sampling: Random selection across multiple customers and industries ensures results reflect real-world diversity.
  • Conservative Scoring: Penalizing both missed and incorrect identifications provides a realistic performance measure rather than inflated metrics.
  • Regular Validation: Bi-weekly testing with manual verification maintains accuracy standards over time.
  • Large Sample Size: 40,000 product instances provide statistically significant, reliable results.

Methodology Assessment​

This is an excellent, conservative approach that exceeds industry standards. The methodology is particularly strong because:

  • The penalty system for false positives prevents inflated accuracy claims.
  • Cross-customer sampling eliminates bias toward specific product types or environments.
  • Regular bi-weekly testing catches performance drift.
  • Manual verification ensures ground truth accuracy.

This conservative scoring method likely produces lower but more trustworthy accuracy figures compared to competitors who might only count missed detections as errors.