Automatic performance evaluation for video text detection

摘要: We propose an objective, comprehensive and difficulty-independent performance evaluation protocol for video text detection algorithms. The includes a positive set negative of indices at textbox level, which evaluate the quality in terms both location accuracy fragmentation detected textboxes. In protocol, we assign difficulty (DD) level to each ground truth textbox. can then be normalized with respect DD are therefore independent difficulty. also importance (DI) overall rate is DI-weighted average qualities all textboxes, makes more accurate reveal real performance. automatic scheme has been applied on approach determine best parameters that yield results.