The new algo caches extra frames so it can select a "best" frame to drop
from a list. A metric for each buffer is calculated as it is added to
the list. The metric indicates how much the frame differes from the
previous frame. The one with the lowest metric is selected for dropping
when a drop is required.
The old algo tried to latch on to a pattern without keeping any extra
frames. When "in sync" it would drop the buffer that fit the pattern.
But this only worked for a few specific patterns I had tested with and
did not work for all possible patterns (e.g. issue 50 where 1 in 5
needed to be dropped to convert 29.97 to 23.976).
projects / handbrake / commit