Repository navigation
Feature Request: error analysis visualization #887
Description
Activity
- added a commit that references this issue
on Sep 23, 2026 Hello @christuchez, thanks for the feature suggestion!
This is merged into dev in #890 and #902 and will ship in the next release (v1.6.0).
Running model.val(data="data.yaml", visualize=True) saves validation images annotated with true positives, false positives, and missed detections. Images with errors go into visualize/errors/. You can also filter them in code using results.box.image_metrics.
Let me know if this covers what you needed
reopen the issue if you follow up pls
- added 2 commits that reference this issue
on Sep 26, 2026
When running validation on the test/validation set, it would be useful to visualize the samples the model got wrong. This should be configurable to show only the first N incorrect samples, to avoid generating too many for large datasets. Each incorrect sample should have overlaid text showing the ground truth and predicted label, with the text sized small enough to fit within the image.
The value of this is enabling error analysis — identifying where the model is failing, spotting systematic errors, and catching mislabeled samples.
For detection tasks you could also show the samples that were missed along with the ones that are the wrong class.