Annotate Faster with Box Suggestions

Correcting a suggested box is much faster than drawing it. With a box suggestion service, LumiScan Annotation Web asks your trained network for boxes on the open image; you only check and correct them.

The loop

  1. Annotate the first few hundred images by hand.

  2. Train a network with LumiScan Training.

  3. Copy the exported network into the box suggestion service (see Box Suggestion Service).

  4. Annotate the next images with suggestions: accept, correct, add what is missing.

  5. Train again with the larger data set. Every round makes the suggestions better.

For parts without a network yet, the service can run an open-vocabulary detector that finds objects by the names of the labels.

1. Choose the network for the project

A project admin opens Settings and, under Box suggestions, chooses the Network of the project. If the service has only one network, it is used without choosing. Also set:

  • Suggest from confidence %: suggestions below are left out (default 30 %).

  • At most boxes per image: default 50.

2. Get suggestions

Open an image and click Suggest boxes (P). After a moment, the suggestions appear as dashed boxes with their confidence:

An image with dashed suggested boxes labelled bracket, nut and washer with their confidence in percent; on the right the note that the suggestions are not accepted yet, with the buttons Accept all and Discard

Boxes already on the image and suggestions outside the region of interest are not suggested again.

3. Accept, correct or discard

  • Correct a suggestion (move, resize, rotate, change the label): it is accepted.

  • Select a suggestion and press Enter: it is accepted as it is.

  • Right-click: it is deleted.

  • Accept all / Discard handle all remaining suggestions.

Seven suggested boxes not accepted yet. Correcting one, or Enter on the selected one, accepts it. Buttons Accept all and Discard

If you save with suggestions still open, you are asked whether to accept them or leave them out. Suggestions are never saved by themselves.

Check every suggestion, also those with a high confidence. A network trained on corrected suggestions repeats its own mistakes if they are not corrected.