Learn & Complete
Learn & Complete is a workflow that speeds up building annotation datasets. Instead of manually labeling every image, you correct a small set, run Learn to train the model on those examples, then run Complete to let the model fill in the annotations on the rest.
This is an index workflow. It works on the images and observations stored in your indexes, not on recorded bundles.
The idea
Building a good training dataset by hand is slow. Learn & Complete shortens the loop:
- Correct a small set of images — draw the right boxes or place the right keypoints.
- Run Learn — the model trains on just those examples.
- Run Complete — the model fills in boxes or keypoints on a larger set.
- Review the results, fix what needs fixing, and repeat.
Each cycle improves the model and grows the dataset. Once you are happy with the annotations, do a full training run to consolidate everything.
Where to find Learn & Complete
There are two entry points, each suited to a different situation.
Selected images
Right-click one or more selected images in the index. Under the ROI or Keypoints submenu you will find Learn and Complete.
Use this when you want to work with a specific hand-picked set of images — for example, a group of images you have just corrected or a set of similar cases you want to annotate together.
The actions are available when:
- at least one image is selected
- the active model supports the annotation type (boxes or keypoints)
- ROI is enabled (for ROI actions) or keypoints are enabled for the category (for keypoint actions)
From a category
Right-click a category card in the index panel. The category menu shows Learn and Complete directly.
Use this when you want to work at the level of a whole category — for example, training on everything already labeled for that category and then finding that same type of object across images from other categories.
The actions are available when the category has at least one observation.
Learn
Learn runs a focused training session using the current model as a starting point. It does not reset the model — it continues from where it is and refines it on the selected scope.
What Learn trains on
- From images: trains on the selected image files. If a file has observations from multiple categories, all of them are included in the training sample so the model keeps its broader knowledge.
- From a category: trains on all image files that belong to that category. Other observations on those same files are also included.
The Learn dialog
| Setting | Default | Description |
|---|---|---|
| Nr of training epochs | 10 | How many passes over the training data |
| Training mode | Heads only | Which parts of the model to train |
| Size of test data (%) | 0% | How much of the selection to hold back for validation |
Training mode
Heads only freezes the backbone and trains only the detection heads. This is faster and safer for small selections — the backbone's general knowledge is preserved, and only the prediction layer adapts.
Full model allows every part of the model to continue training. Use this when your selection is representative enough that you want deeper adaptation.
What Learn changes
Learn starts from the current model configuration and applies focused overrides:
- training scope is restricted to the selected files
- learning-rate decay uses cosine
- patience is set to zero — no early-stopping wait
- the model continues from its current weights, not from scratch
Learn is not supported for identify models.
After using Learn & Complete to build up your annotations, always do a full training run to properly consolidate the dataset into the model.
Complete
Complete runs the current model over a set of images and saves the boxes or keypoints it finds. The submenu you used to open Complete determines what gets written — ROI Complete writes boxes, Keypoints Complete writes keypoints.
Complete from images
The Complete dialog asks for:
| Setting | Default | Description |
|---|---|---|
| Min score | 0 | Minimum confidence for a detection to be kept |
| Max nr of detections | 1 | Maximum number of results saved per image |
| Remove existing first | Off | Whether to clear existing annotations before saving new ones |
After you confirm, Complete processes all selected images using the model and saves results directly. You can continue working while it runs.
Complete captures the active index and category at the moment you start. Results are always written for that index and category, even if you navigate elsewhere while it runs.
Complete supports images only. If the selection includes video observations the operation stops and reports the issue.
Complete from a category
When started from the category menu, Complete works differently. Instead of processing your selection, it looks across the other categories in the active index for images that do not yet belong to the category you right-clicked.
It runs the model on those images and collects any detections for your category. Before anything is saved, a review dialog opens showing all the proposed results. You pick which ones to keep and discard the rest, then confirm. Only the items you accept are saved.
This is useful when a new category is being added and you want to find all the places it already appears in your existing footage — without manually searching through every image.
Complete from a category does not have a "remove existing" option. It only creates or updates observations for the target category and never touches annotations belonging to other categories.
ROI vs keypoints
The submenu you open determines what Learn & Complete works on.
- Start from ROI → the workflow is about bounding boxes.
- Start from Keypoints → the workflow is about keypoints.
If your model supports both, the two are independent. Completing boxes does not fill keypoints, and completing keypoints does not fill boxes.
Tips
Start small and clean. Learn produces better results when the selected examples are correct. A small set of well-labeled images teaches more than a large set with mistakes.
Use filtering options to build your selection. Filter by category, attribute value, or selection state to find the images that most need annotation, then correct and Learn on those.
Review Complete results before training. Complete is a fast annotation assistant, not a final step. Treat its output as a first pass and correct anything that looks wrong before using the data for a full training run.
Iterate. A single Learn & Complete cycle rarely finishes the job. Each pass improves the model a little, which makes the next Complete more accurate. A few short cycles often work better than one big one.
For model setup and training settings, see Model Configuration.