Confidence, Uncertainty, and Human Review
A detection marked 0.9 can look reassuring, but that number is not automatically a 90 percent guarantee of correctness.
Confidence, Uncertainty, and Human Review explained visually
A detection marked 0.9 can look reassuring, but that number is not automatically a 90 percent guarantee of correctness. English captions are optional and start off. The complete transcript follows the written lesson.
What you will learn
A detection marked 0.9 can look reassuring, but that number is not automatically a 90 percent guarantee of correctness.
A score is not a promise
A detection marked zero point nine can look reassuring, but that number is not automatically a ninety percent guarantee of correctness. Confidence scores depend on the model and training procedure. Their relationship to actual correctness must be checked on representative data. Calibration asks whether groups of predictions with similar scores behave as those scores suggest.
Evaluate unfamiliar conditions
Unfamiliar inputs create another difficulty. A camera, weather condition, or object type outside the training distribution can produce unreliable predictions, sometimes with high confidence. Test these conditions explicitly. An out-of-distribution detector or uncertainty estimate can provide useful signals, but it also has limitations and needs evaluation.
Design useful review
Human review works best when the interface supports correction without hiding the original evidence. Keep generated annotations editable, preserve manual corrections, and distinguish observations from estimates. Decide which cases deserve extra attention, such as low overlap, large jumps, or crowded crossings. Review is not merely a warning message; it is a workflow that helps people make accurate decisions efficiently.
A worked example
If one hundred predictions all receive scores near zero point eight and only fifty-five are correct under the evaluation rule, those scores are overconfident for that group. Calibration should be assessed on held-out, representative examples rather than inferred from one attractive prediction.
Worked example
If 100 predictions all receive scores near 0.8 and only 55 are correct under the evaluation rule, those scores are overconfident for that group. Calibration should be assessed on held-out, representative examples rather than inferred from one attractive prediction.
Quick check: test your understanding
Try it yourself: Name two inputs that might challenge a model trained only on clear daytime images.
Check the reasoning
Night scenes and heavy rain are two examples. Other changes include infrared imagery, unusual viewpoints, or objects much smaller than those in training.
Should a high confidence score override a clearly incorrect box seen by a reviewer?
No. The reviewer should correct the annotation based on the image evidence and labeling rules.
How would you solve the practice case?
Night scenes and heavy rain are two examples. Other changes include infrared imagery, unusual viewpoints, or objects much smaller than those in training.
What does the worked example show?
If 100 predictions all receive scores near 0.8 and only 55 are correct under the evaluation rule, those scores are overconfident for that group. Calibration should be assessed on held-out, representative examples rather than inferred from one attractive prediction.
Further reading
Programming Computer Vision with Python — Jan Erik Solem.
Original ToolsInVision teaching examples. Editorial approach · Suggest a correction
Accessible referenceLesson 14 transcript202 words · Video transcript · Matches the published narration and captionsOpen transcript
A detection marked zero point nine can look reassuring, but that number is not automatically a ninety percent guarantee of correctness. Confidence scores depend on the model and training procedure. Their relationship to actual correctness must be checked on representative data. Calibration asks whether groups of predictions with similar scores behave as those scores suggest.
Unfamiliar inputs create another difficulty. A camera, weather condition, or object type outside the training distribution can produce unreliable predictions, sometimes with high confidence. Test these conditions explicitly. An out-of-distribution detector or uncertainty estimate can provide useful signals, but it also has limitations and needs evaluation.
Human review works best when the interface supports correction without hiding the original evidence. Keep generated annotations editable, preserve manual corrections, and distinguish observations from estimates. Decide which cases deserve extra attention, such as low overlap, large jumps, or crowded crossings. Review is not merely a warning message; it is a workflow that helps people make accurate decisions efficiently.
If one hundred predictions all receive scores near zero point eight and only fifty-five are correct under the evaluation rule, those scores are overconfident for that group. Calibration should be assessed on held-out, representative examples rather than inferred from one attractive prediction.