Tracking Error Correction Mode v2.5

Use Case

You track some frames in a project. As you scroll through the results, you take note of frames where the tracker performs poorly. You would like to add these "bad frames" to the training data, with manually corrected labels, so that the tracker can learn and improve its performance on similar frames.

A related use case occurs when it is not CPR, but an external tracker that generates the tracking results. In this case, the only difference is that i) the tracking results have been generated externally and imported to your project, and ii) any labels you create will be exported back to the tracker for retraining. Otherwise, the main browse-and-correct workflow described below is the same.

Functional Spec

1. Frames are tracked in a project, or tracking results are imported.
2. While browsing the results, poor tracking performance is noted and you decide to augment the training data near/at bad frames.
3. Switch labeling mode into "Error Correction Mode".
4. Error Correction Mode resembles Template mode, except that the template is taken from the predictions rather than the previous frame/labels:
   • When present, manual labels are shown by themselves. (By default in this mode, Hide Predictions is turned on; toggle Hide Predictions to get a glimpse of tracking results.)
   • When manual labels do not exist but tracked predictions do, the tracked predictions are shown and used as "Template" points (resembling tracking points but eg faded)
   • When neither manual labels nor tracked predictions exist, either no points are shown, or the "Template" points are shown in white.
5. While browsing frames, the template/prediction points allow you to note frames with poor tracking. When desired, you can adjust these points just as in Template mode. Hitting Accept creates a new labeled frame.
6. When you are done, you will have added to your project a (typically, sparse) set of manual labels at selected frames where the tracker performed particularly poorly.