[FFmpeg-devel] [PATCH] lavfi: port mcdeint filter from libmpcodecs

[Bryce Harrington]

On Tue, May 28, 2013 at 03:40:43PM +0200, Stefano Sabatini wrote:
> On date Monday 2013-05-27 19:43:01 +0200, Stefano Sabatini encoded:
> [...]
> +/**
> + * @file
> + * Motion Compensation Deinterlacer
> + * Ported from MPlayer libmpcodecs/vf_mcdeint.c.
> + *
> + * Known Issues:
> + *
> + * The motion estimation is somewhat at the mercy of the input, if the
> + * input frames are created purely based on spatial interpolation then
> + * for example a thin black line or another random and not
> + * interpolateable pattern will cause problems.
> + * Note: completly ignoring the "unavailable" lines during motion

typo:  completly -> completely

> + * estimation did not look any better, so the most obvious solution
> + * would be to improve tfields or penalize problematic motion vectors.
> + *
> + * If non iterative ME is used then snow currently ignores the OBMC
> + * window and as a result sometimes creates artifacts.
> + *
> + * Only past frames are used, we should ideally use future frames too,
> + * something like filtering the whole movie in forward and then
> + * backward direction seems like a interesting idea but the current
> + * filter framework is FAR from supporting such things.
> + *
> + * Combining the motion compensated image with the input image also is
> + * not as trivial as it seems, simple blindly taking even lines from
> + * one and odd ones from the other does not work at all as ME/MC
> + * sometimes has nothing in the previous frames which matches the
> + * current. The current algorithm has been found by trial and error
> + * and almost certainly can be improved...
> + */
> +