Local dataset quality check
Find invalid bounding boxes
Boxes can be empty, inverted, out of bounds, tiny, extreme, overlapping, or duplicated.
What to check
Review geometry findings and the affected annotations before a model sees them.
Why it matters
Bad geometry teaches a detector the wrong target and can make model experiments misleading.
When Pro helps
Pro unlocks bulk corrections and saved repair sessions.
What your report can look like
Bounding-box review
An illustrative example of the findings and repair guidance Labelyst produces for your dataset.
Status: Blocked
Score: 62/100
2 findings
Findings
- 3 boxes have inverted coordinates
- 1 box has zero area
Passed check
- Image dimensions were available for checked boxes
Suggested next step: Correct invalid geometry before training.
Rules: standard profile · local check
Illustrative local result. It is not a certification.
Check it before training
No account or upload is required for a Free check. Analysis, repairs, and reports stay on your device.