Multi Target Tracking

Open Issues

• For the moment, the following features are unavailable for multi-target projects (projects with trxfiles):
  • Histogram equalization
  • "Full" tracking results with replicate cloud visualization
  • Incremental Training (Full Retrain only)
  • ... probably others as well
• Currently the trx orientation (.theta) is not used for anything. Typically you will want rotCorrection on (as well as Replicates>Randomize initial orientations).
• Currently, all targets are tracked during any Tracking operation.

Labeling

Label frames/targets as you normally would. All labeled targets/frames will be used for training.

CPR Parameters

There are two new tracking parameters related to multi-target tracking:

• Target cropping>Target crop radius. For training and tracking, image patches are cropped from the movie using a square ROI centered at the target trajectory. This parameter sets the radius (one-half the side length) of this square ROI. The center of the ROI is defined to be (trx.x,trx.y) for the target trajectory. Enhancement: allow guided/intelligent guesstimate.
• Target cropping>Pad background. Scalar grayscale "background" value (in [0,1]) for padding patches/ROIs that fall outside the video frame. When a target ventures near the edge of its video, its square patch/ROI may fall outside the video frame. Pixels in the ROI that fall outside the video in this manner have their values set to this constant grayscale value. Possible enhancement: maybe this could be computed as a mean or median pixel intensity instead.
• When targets are free to wander in an open arena, their heading (rotational orientation) will be arbitrary. In this situation, RotationalCorrection>Enable should be checked, and landmarks should be chosen for the Head and Tail.
  • New 20171004: For purposes of determining the orientation of a shape, the precise subset of points in the range [Head,Tail] is used. Landmarks/points outside of this range are considered satellites/limbs and do not impact a shape/animal's orientation. This is a technical point, but of some practical importance. Points in the range Head..Tail are used to extract/compute an overall orientation for the animal and therefore these points should be defined to be stable "body" points.

Training Montage

After training, Track>Training Data Montage lets you view a randomly-generated montage of ROI patches created/used for training. Enhancement: Don't require a full train.