Is it the sparsity or collaborativeness that makes a visual tracker strong?

Previous work has developed a visual tracking algorithm, based on sparsity, that represents a target as a superposition of templates from a gallery in a fashion that the coefficients are sparsely populated. When occlusions occur, sparsity is maintained by bringing additional trivial templates (identity bases) into that gallery. While reported desirable results in visual tracking applications, several researches in recognition community have questioned the effectiveness of imposing L1 norm based sparsityconstraint and recommended collaborative representation, which replaces L2 norm as the measure of sparsity. Little work has been done in visual tracking to access the usefulness of the sparsity for visual tracking. This work aims to present a study on sparse and collaborative representation in the context of visual tracking and demonstrate which representation is really useful to achieve better tracking performance. To this end, extensive experiments are conducted on several challenging sequences and a discussion based on the experimental comparison is presented.