Visual Object Tracking Performance Measures Revisited

摘要: The problem of visual tracking evaluation is sporting a large variety performance measures, and largely suffers from lack consensus about which measures should be used in experiments. This makes the cross-paper tracker comparison difficult. Furthermore, as some may less effective than others, results skewed or biased toward particular aspects. In this paper, we revisit popular visualizations analyze them theoretically experimentally. We show that several are equivalent point information they provide for and, crucially, more brittle others. Based on our analysis, narrow down set potential to only two complementary ones, describing accuracy robustness, thus pushing homogenization methodology. These can intuitively interpreted visualized have been employed by recent object challenges foundation