[FFmpeg-cvslog] r19923 - in trunk/tests: codec-regression.sh rotozoom.regression.ref vsynth.regression.ref

Reimar Döffinger Reimar.Doeffinger
Sat Sep 19 22:47:26 CEST 2009


On Sat, Sep 19, 2009 at 11:51:15AM -0700, Baptiste Coudurier wrote:
> >  14400000 ./tests/data/a-dv50.dv
> >  a2ff093e93ffed10f730fa21df02fc50 *./tests/data/dv50.vsynth.out.yuv
> >  stddev:    1.72 PSNR: 43.37 bytes:  7603200/  7603200
> >-78ae79c285b3a7607b8be96e925d0795 *./tests/data/a-dnxhd-1080i.mov
> >+8ff3c18dabe28509b4e5b3fdaa7f0856 *./tests/data/a-dnxhd-1080i.mov
> >  30311415 ./tests/data/a-dnxhd-1080i.mov
> >-6674aca7d4bc08543a3c316af82207c0 *./tests/data/dnxhd.vsynth.out.yuv
> >-stddev:    2.24 PSNR: 41.10 bytes:  7603200/  7603200
> >-fcb138df8e991e2e2133044da234f36b *./tests/data/a-dnxhd-720p.dnxhd
> >+52f50a8031eebf980a1303e4744a49f8 *./tests/data/dnxhd.vsynth.out.yuv
> >+stddev:    6.28 PSNR: 32.16 bytes:  7603200/  7603200
> >+d206daba40f3c1be8b834de8a41e984c *./tests/data/a-dnxhd-720p.dnxhd
> >  22937600 ./tests/data/a-dnxhd-720p.dnxhd
> >-a5810fbc20aaed082b701c839da10c75 *./tests/data/dnxhd.vsynth.out.yuv
> >-stddev:    1.28 PSNR: 45.97 bytes:  7603200/  7603200
> >+d7f14ee0dd3a4df95f05999677da348b *./tests/data/dnxhd.vsynth.out.yuv
> >+stddev:    6.39 PSNR: 32.00 bytes:  7603200/  7603200
> >+c0a42b916489d3c1d0a315974821e28d *./tests/data/a-dnxhd-720p-rd.dnxhd
> >+22937600 ./tests/data/a-dnxhd-720p-rd.dnxhd
> >+50efe763b272af3367985218aff68b1b *./tests/data/dnxhd.vsynth.out.yuv
> >+stddev:    6.25 PSNR: 32.20 bytes:  7603200/  7603200
> >  595fc4e38734521356b60e67b813f0fa *./tests/data/a-svq1.mov
> >  1334367 ./tests/data/a-svq1.mov
> >  9cc35c54b2c77d36bd7e308b393c1f81 *./tests/data/svq1.vsynth.out.yuv
> 
> Is it me or -sws_flags neighbor+bitexact produce a lot better output ?

Depends on what you mean by "better", but PSNR is better for nearest neighbor
when first up- and then downscaling.
Still I think it is a bad idea to hack this in on a per-test basis, also using
nearest neighbor for downscaling after decoding means the decoder could break
about 3 of 4 pixels and the output wouldn't change.



More information about the ffmpeg-cvslog mailing list