Skip to content

Draft: Filter training data for face recognition models

Philip Gouverneur requested to merge gouvi/digikam:filter_face_training into master

As has been pointed out in various issues, face recognition is not always accurate. Especially when a large number of faces have already been confirmed, the accuracy leaves a lot to be desired. There are also many ideas on how to improve accuracy by limiting the training data of the model. Therefore, this merge request introduces the idea of limiting the training data in the "Maintenance" tool, for example by removing images older than a certain date. More (complex) ideas could also be included here.

Related issues:

449310 437896 485029 425981

Draft: This merge request is currently a draft, for the following reasons:

  • Currently, the approach to limit images by time uses the "modification date" but should use the "creation date" of the image
  • "Last 100 images per person" is not implemented so far
  • Other filters could be implemented
  • I wanted to check with you if I should follow this idea before finalizing it

Thank you for your consideration and the hard work!

Merge request reports