Skip to content

Labeling and Predicting Occlusions

roian edited this page Jan 13, 2020 · 2 revisions

For points in the frame, but occluded by another body part, use the shift-click or select point 'o' method to create a partially occluded label estimating where the point is located.

For points that are out of frame, or where you can't make a reasonable estimate of the location of an occluded point, use the fully-occluded label (only works with MDN currently). To use fully-occluded labels, select View>Show occluded points box - advanced to toggle on the fully-occluded box. (A weird bug might still exist where if you change labeling methods, the box has to be toggled off-on again to work.)

Under advanced tracking parameters, the following options need to be checked: Ignore occluded landmarks, predict occluded.

When labeling points, click inside the fully-occluded box to mark a point as fully-occluded.

In the trk file the variable pTrkocc is the occlusion statistic, e.g. a float/double in [0,1]; threshold to 0.5 for a binary yes/no predicted-as-occluded flag.