Error analysis for vision
Systematically inspecting vision mistakes by slice: lighting, size, background, demographics, camera type—then fixing data/model/threshold.
What it is
Systematically inspecting vision mistakes by slice: lighting, size, background, demographics, camera type—then fixing data/model/threshold.
Why it matters
Average mAP hides who/what fails. Error analysis is how products improve safely.
How it works (plain)
Sample failures → tag causes → quantify tags → prioritize fixes → re-eval slices—not only globals.
Try it
Take 20 wrong detections; invent a tag taxonomy of at least 5 causes.
Myths
- ⚠️ Myth: More epochs fix all error tags.
- ✓ Reality: Many errors are data/label/threshold issues.
Sources
- Course 12 detection/bias; Course 18 evaluation
- fast.ai practical tips: https://www.fast.ai/ ↗