LabelBOT

[mzur]

This is about the implementation of LabelBOT. The details have already been discussed. Below is my mockup of the UI. Further discussions can happen here.

[mzur]

Maybe this will finally require a merge of biigle/largo into biigle/core (#362). otherwise the required feature vector tables will not be available. Largo is installed in all default BIIGLE instances anyway.

[mzur]

Here are some thoughts I just found in my notes:

• When the random/regular sampling mode is active, LabelBOT should immediately suggest a label for the currently active point. When the label is confirmed with Enter or a click, the annotation mode advances to the next point. The LabelBOT timeout is disabled here, i.e. users always have to confirm the chosen labels. The suggested label for the next point could even be prefetched while the user is still looking at the current point.

• There could be an integrated random/regular sampling and classification workflow at some point. This is started like the laser point detection. It computes random/regular point locations and for each runs LabelBOT to get a label. When the processing is finished the user is notified. The user then can review these using Largo.

  Alternatively we could change MAIA into kind of a general purpose review tool for machine-generated annotations. Users could review automatic object detection (w/ or w/o classification), point to polygon conversion or random/regular sampling w/ classification results there.

• There is a limit of up to 5 pending LabelBOT requests per user. This limit can be tracked in a cache variable like the video tracking. Requests above this limit are rejected but the UI should handle this gracefully and tell users about this instead of sending actual requests that produce the rejection.

• If computing the feature vector client-side takes a significant amount of time, we could try to compute vectors at fixed intervals already while the user draws the annotation. If the final annotation is reasonably contained in the vector bounding box that was previously computed, immediately send the vector to the backend instead of computing it only after the annotation was finished.

• LabelBOT will first be implemented for images then for videos.