Perception-Oriented Methodology for Robust Motion Estimation Design

Optimizing a motion estimator (ME) for picture rate conversion is challenging. This is because there are many types of MEs and, within each type, many parameters, which makes subjective assessment of all the alternatives impractical. To solve this problem, we propose an automatic design methodology that provides `well-performing MEs' from the multitude of options. Moreover, we prove that applying this methodology results in subjectively pleasing quality of the upconverted video, even while our objective performance metrics are necessarily suboptimal. This proof involved a user rating of 93 MEs in 3 video sequences. The 93 MEs were systematically selected from a total of 7000 ME alternatives. The proposed methodology may provide an inspiration for similar tough multi-dimensional optimization tasks with unreliable metrics .