Trk File: exported predictions or labels - kristinbranson/APT GitHub Wiki

MATLAB .mat file with .trk extension

Can be loaded into MATLAB using load(trkfilename,'-mat')__

Exported labels have _labels at the end of automatically generated file name.

Contains following variables:

pTrk (numlandmarks x numdimensions x numframes x numtargets) - Contains the position information for tracked points (or labels), unlabeled frames have NaN values. For example: if there are two labeled body points for a single target from a single viewpoint video with 10800 frames, then pTrk is 2x2x10800, where pTrk(1,1,15) is the x coordinate of the first point for frame 15, pTrk(1,2,10400) is the y coordinate of the first point for frame 10400. First dim of pTrk is labeled by pTrkiPt, the 3rd dim is labeled by pTrkFrm, and the 4th dim by pTrkiTgt

pTrkTS (numlandmarks x numframes x numtargets) - TimeStamp at which the frames were labelled (or tracked) (-Inf for unlabeled)

pTrkTag (numlandmarks x numframes x numtargets) Occlusion data, 0=unoccluded and is the default, 1==occluded and is what you get using shift-click in Sequential labeling mode or the letter o when a point is selected in Template mode.

pTrkFrm (1 x numframes) These are the frames labeling the 3rd dimension of pTrk (and 4th dimension of pTrkFull). In some cases a TrkFile can be generated where pTrk doesn't cover the whole movie, only a subset, in which case pTrkFrm tells you which frames are covered. The usual case though is that .pTrk covers the entire movie and then .pTrkFrm will be 1:totalNumFrames. Lists all frames in pTrk file including those without tracking data.

pTrkconf for DL trackers there are confidences reported for the tracking results. What is saved varies on output of tracker.

pTrklocs_DLtracker intermediate tracking locations, final saved in pTrk.

trkInfo Information about the tracker used, if applicable. Empty in the case of manual labels.
fields: paramFile, param

Only in CPR trk files:

pTrkFull (numpoints x numdimensions x numreps x ~numframes) numreps: the number of CPR 'replicates' or number of 'test shapes' that are propagated using the trained regressor cascade. Some central measure of this cloud of shapes is taken to be the single tracking result. In the param.yaml, this is TestInig:Nrep.

pTrkFullFT (size of 4thDimpTrkFull x 2) Frame-Target table labeling 4th dim of pTrkFull.