Motion vector recovery method based on kernel regression

In this paper, we describe a motion vector (MV) recovery method using kernel regression. In H.264, MVs are assigned based on 4×4 block, which means that MVs contained in neighboring macroblocks (MB) are highly correlated. Taking those available MVs as training set, recovery of lost MV can be considered as regression in local MV field. Then a nonparametric kernel smoother is applied to estimate lost MVs in the corrupted MB. The associated bandwidth estimation is derived by analyzing standard deviation statistically. Experimental results show their better performance.