A feature tracking algorithm using both forward and backward correspondences is described. The algorithm can be divided into three stages: initialization, tracking, and postprocessing. The initialization stage takes the first three frames and finds a set of best matches in terms of some cost function. The tracking stage extends trajectories for each new frame based on the same cost function. The postprocessing takes care of linking fragmented trajectories by extending the search area. Thus the algorithm can handle occluded, entering, and exiting objects.

The major question I have is about the convergence of the initialization step. Evidently, the initial configuration is dependent on the processing order of points. The authors claim that two iterations is enough to reach a near-optimal configuration with point sets of “reasonable” density. What is “reasonable,” and does the algorithm break down at a certain density? The authors do not provide clear answers to these questions.

Their experiments are not extensive but, according to the authors, detailed performance evaluation is reported elsewhere. The references are not complete, and some important works (in particular, Kalman filter–based techniques) are not mentioned. Overall, the paper is readable and easy to follow.