Failure review
Compare prediction and human truth for each image and record what went wrong.
Before you start
Select available dataset, split and model. The chosen split may be empty.
Quick start
- Choose an image and inspect Original, Overlay or Mask views where available. Toggle Human revision to compare. Click the image list and use Up/Down to browse.
- Describe missed, extra, merged or wrongly classified objects and their context. Save the review and choose a route.
- Open images needing correction in Annotator.
Functions and choices
Find the right images
Use search, dataset, split and review filters to narrow the list. Click the image list and use Up/Down to browse.
Run AI prediction
AI prediction runs the selected model on all filtered images. The count shows the batch size; progress shows processed images and errors. Stop cancels the remaining work. Human annotations stay unchanged.
Compare annotations
Solid lines show human annotations; dashed lines show model suggestions. Both use project class colors. Show or hide each layer independently. These switches change only the view, not saved data.
Load saved evidence
Import from latest training loads saved results. For images sent from Evaluation, choose Training evaluation and All project images to find every image. Image badges show origin, dataset split and review route; images outside the dataset are labelled.
Choose a route
Quick actions offer OK and five routes: Training, Guard Set, Save Later, Hard Case and Exclude. Guard Set reserves regression examples. Route changes stay in the draft until you save.
Save the review
Save Review or Command-S saves the review and route together. Wait for Saved. If saving fails, the draft remains for retry. Saving does not retrain the model.
Add context
Expand Notes to read or edit your comments. Model version and Test source appear only when saved values are available.
Result
A saved review and routing decision for the evidence.
Next step
Use Annotator for corrections, Data Studio for a new version and Training for the next run.