Skip to main content

Learn & Complete

Learn & Complete is a workflow for improving an index without switching away from the search results.

  • Use Learn when you have corrected a small set of ROI (region of interest) boxes or keypoints and want the current model to learn from that focused selection.
  • Use Complete when you want the app to run the current model on another selected set of image observations and fill in the matching ROI boxes or keypoints.

This is an index and training workflow. It is separate from bundles, which are used for recorded video review and analysis.

When to use it​

Use Learn & Complete when you are building or cleaning a training dataset and want to avoid manually annotating every similar image.

A common workflow is:

  1. Select a small group of images.
  2. Correct their ROI boxes or keypoints manually.
  3. Run Learn on that selected group.
  4. Select another set of images.
  5. Run Complete to fill in ROI boxes or keypoints from the trained model.
  6. Review the generated annotations and correct anything that still needs work.

Where to find it​

Open the context menu on an image search result, then open either:

  • ROI
  • Keypoints

Each submenu can show:

  • Learn
  • Complete

The actions are available only when:

  • at least one search result item is selected
  • the active model supports the annotation type
  • ROI is enabled for ROI actions
  • keypoints are enabled and allowed for the selected category for keypoint actions

Learn​

Learn starts a focused training run for the current model.

It uses the selected search-result files as the training scope. When a selected file contains multiple observations, Vidsy.ai includes the observations on that selected file so related category data can stay part of the sample.

The Learn dialog asks for:

  • Nr of training epochs
  • Training mode
  • Size of test data (%)

The default values are:

  • 10 epochs
  • Heads only
  • 0% test data

Training mode​

Heads only keeps the backbone frozen for the focused run and trains the prediction heads.

Full model allows the full model to continue training during the focused run.

What changes during Learn​

Learn starts from the active model training configuration, then applies focused-run overrides:

  • training uses the selected search-result scope
  • learning-rate decay is set to cosine
  • epochs is set to the dialog value
  • patience is set to 0, so there is no early-stop waiting
  • test-data size comes from the dialog
  • the model continues from the current model instead of resetting
warning

Intermediate Learn is not currently supported for identify models.

Complete​

Complete runs the current model over selected images and saves the boxes or keypoints it finds for each image.

Use ROI Complete to find ROI boxes. Use Keypoints Complete to find keypoints.

The Complete dialog asks for:

  • Min score
  • Max nr of detections
  • Remove existing first

The default values are:

  • 0 minimum score
  • 1 detection
  • remove existing off

What Complete processes​

Complete currently supports images only. If the selection contains video observations,the app stops the operation and reports the problem.

When Complete starts, Vidsy.ai snapshots the active index, active category, selected observations, and selected image files. You can continue working while it runs, but the operation still writes results for the index and category that were active when it started.

How detected results are used​

Complete runs inference through the existing pipeline, then filters the detector output inside the Electron app.

Results are only kept when they:

  • belong to the active category captured at the start
  • meet the minimum score
  • fit within the max-detections limit for the selected image

ROI vs keypoints​

Learn & Complete always follows the submenu you choose.

If you start from ROI, the workflow is about boxes. If you start from Keypoints, the workflow is about keypoints.

This matters when your model supports both. Completing ROI boxes does not fill keypoints, and completing keypoints does not fill ROI boxes.

Practical tips​

Start with a small, carefully corrected selection. Learn is most useful when the selected examples are clean enough to teach the current model the pattern you want.

After running Complete, review the generated boxes or keypoints before using the data for broader training. Treat Complete as a fast annotation assistant, not as a final validation step.

For model setup and training settings, see Model Configuration.

tip

Intermediate - fast learning skews the model. Best to always retrain the full model again after annotations are done.