Přepis řeči - IMPROVED P-DOMAIN RATE CONTROL WITH ACCURATE HEADER SIZE ESTIMATION

0:00:13	one introduction
0:00:14	and uh i had a rabbit
0:00:16	uh today i'm going to uh present um uh improve the meta control always accurate header size estimation
0:00:22	and that this is not one with my adviser professor a constant
0:00:28	so we know that that control is a uh
0:00:31	essential part of a practical video encoder
0:00:35	and uh uh sometimes we care about it a bit rate of the encoded bit-stream
0:00:40	but uh you most uh
0:00:42	a real-time applications we want to control the size of each frame actually to be
0:00:46	and the limited to control it started control method originally proposed
0:00:50	by just stopped at six three and i
0:00:54	uh
0:00:55	the basic assumption you don't need any control is that uh this time frame
0:01:00	proportional
0:01:01	to the percentage of one of nonzero coefficients
0:01:05	uh
0:01:06	in uh after quantization in the frame
0:01:08	uh and experimental results have shown that uh uh automated control can
0:01:14	uh controls the size of the
0:01:16	uh offerings
0:01:17	oh accurately
0:01:19	and they're better than we want to use the mutual matched up
0:01:24	uh the eucharist amount of header information becomes one of the
0:01:29	uh
0:01:30	this is because this side information is not of interest in the original domain into model
0:01:36	so uh if we want to make me to control my to it
0:01:40	well matched up to six
0:01:42	we need to find a way to
0:01:44	uh estimate the size of the information
0:01:47	in the picture
0:01:49	this is our motivation
0:01:52	and uh this is my talk today force this motivation we never talk about
0:01:57	and the next we will introduce some for the header information is
0:02:01	and B also introduce a
0:02:02	two stage with model
0:02:04	to to stage of it can control them based on the mean control
0:02:07	and that this is followed by some experiment with that's
0:02:10	we compare our rate control method space
0:02:13	some
0:02:14	a previous work
0:02:16	and uh family we will draw the conclusion
0:02:20	so
0:02:21	slice you what is included in the header information you
0:02:26	so we know the two uh most common header C natural science writers and a macroblock header
0:02:33	and that the size that it is relatively stable in a
0:02:38	uh uh but once you know the net is in the pudding
0:02:43	and we can accurately estimate the size of the i-th either
0:02:47	and the format headers
0:02:48	so that's not a macroblock header changes
0:02:51	from michael michael and it is harder to estimate
0:02:55	uh
0:02:56	and the industrial for intra macroblocks
0:02:59	yeah because the number of intra macroblocks are quite limited in the people
0:03:04	uh uh we simply use L O
0:03:08	had a size of the encoded in intra macroblocks
0:03:11	to estimate the uh had a size of intra macroblocks in the current frame
0:03:16	and for uh
0:03:17	inter macroblocks we see that there so that different
0:03:21	uh had informations in a in the head inter macroblock
0:03:25	first is the managing information including the motion vectors and the reference frame ideas
0:03:30	and that is the timing information
0:03:32	uh it's a macroblock types and test
0:03:35	and that's there is this information so the code and sir isaac information
0:03:42	yeah and then this figure shows the percentage of
0:03:45	uh different uh
0:03:47	uh kind of
0:03:48	how the information
0:03:49	in relation to the
0:03:51	uh per head aside enough
0:03:53	so the x-axis here this is the frame index and basically this is the percentage of different information
0:04:00	so that the name
0:04:02	oh really
0:04:04	uh is relatively small
0:04:06	so we simply use the information in the past to estimate the
0:04:10	uh and then
0:04:12	up to now should be it's for michael in the current frame
0:04:16	and the first time information
0:04:18	you know the are only the data macroblock types and the block i
0:04:22	a well beyond a need some counter and that just type information is encoded using a a a uh using
0:04:26	a a a a a a fixed to be as a table
0:04:29	so a to be a a the number of different kinds of
0:04:33	uh a block and rocks
0:04:35	you can uh estimate uh the size for
0:04:38	uh a information X to D
0:04:41	and the form shape information and uh as the information be see that the change at a a problem for
0:04:46	inter N
0:04:47	and uh uh the also a a a a to apply a large part of the total had a size
0:04:53	so in the family we real focus on uh the estimation of the stats for
0:04:57	motion information and a
0:04:59	uh us to P information
0:05:03	a
0:05:04	let's
0:05:04	yeah and of the a a motion information so in two N seven
0:05:08	uh fashion
0:05:09	so she's is that a a student of professor clear
0:05:12	and the shouldn't has a a uh uh uh proposed a a a a good of a model for the
0:05:17	uh most information
0:05:19	so uh
0:05:21	does more or she the that uh
0:05:24	so i of the motion information based uh statistics up the
0:05:28	um motion vectors
0:05:29	so is you just had and at this is the number of
0:05:32	not to a motion vector element
0:05:34	and that motion vector element is simply add the the the radical
0:05:38	component
0:05:39	uh a from of the map
0:05:41	and uh this is a a a a a a a number of
0:05:44	this is number of motion actors
0:05:46	and uh uh on the is that a constant is that we can derive from the experiment
0:05:51	and that a i is that a a problem to a we can uh we need to update from frame
0:05:56	to five
0:05:57	and the experiment or that's have shown that that this man X back or many it has a this
0:06:03	uh but but in our experiment we find that it is more of uh that's not to
0:06:08	or is pretty the resulting bit-rate rate
0:06:11	the resulting
0:06:12	uh a number of bits actually be
0:06:14	so uh C here are being called the different uh a sequence it is at a given the bit rate
0:06:19	and that's access
0:06:20	that's a X X six Q a
0:06:22	this is uh a just this i in the bracket
0:06:25	and the back
0:06:26	here this this is the
0:06:27	uh and number
0:06:29	a a and the red points here we show the number of
0:06:33	notion bits
0:06:34	inside a a a a frame
0:06:37	uh
0:06:37	so uh
0:06:39	this is that's in are uh
0:06:41	so the to is not very uh
0:06:44	on linear relationship between the number of of and bits and of this that them here
0:06:48	so we think the problem here is that a a a a uh i a to is for uh that's
0:06:53	not directly in as the motion vector
0:06:56	instead that it to use put coding and the down is a
0:06:59	difference motion vector
0:07:01	so think that a and number of most of bits should have a strong relationship
0:07:05	with the statistics of the difference motion back
0:07:09	yeah so based on this idea
0:07:11	so we that we slightly modify to small
0:07:14	a a for a better prediction
0:07:16	so this is it is just is our modified model
0:07:19	is see that the not we use that uh and number of nonzero
0:07:23	difference motion to elements
0:07:25	and that this is the number of two
0:07:28	uh a difference motion vector elements
0:07:30	and then again i mean that is a a constant we can derive from the experiment
0:07:35	and uh and uh a a a uh we need to apply for mac to mac well
0:07:39	so i we have you down the
0:07:41	uh a in but the experiments for the same has to use this
0:07:45	we see that an hour
0:07:47	uh a a and so this X X this is this have mean the bracket
0:07:51	and the by its is it's is a number of motion bit
0:07:55	we see that an out there is a a a a a linear relationship
0:08:00	between between the number of
0:08:02	um motion bits and this item here
0:08:04	so we think this model can be used
0:08:06	to uh pretty the side of a motion information that
0:08:11	uh and uh
0:08:12	uh but if we look at the this group point thought the was a number of header bits in a
0:08:18	frame
0:08:18	so if we look at the uh uh uh point we see that there is
0:08:22	no a strong relationship between to
0:08:24	the a of number of that the a number of bits and that this item here
0:08:28	so as we have mentioned
0:08:29	uh uh another very and the part and the macroblock block had a size is the
0:08:34	a that information
0:08:37	uh to
0:08:38	that the add to a to estimate the size
0:08:40	a a a a a a a of the a a had information be also need a
0:08:44	man
0:08:45	to uh uh at the me the side of the information
0:08:50	uh you paper speaker show she uh she has proposed that uh the number
0:08:55	all of us to be it's should be proportional to the
0:08:59	uh
0:09:00	uh a number of texture bits
0:09:02	and the been a from the mean that the model of that's and number of texture bits
0:09:06	sure that be a proportional
0:09:08	to the percentage of nonzero coefficient
0:09:11	so here we simply down some experiment is that uh
0:09:13	a uh we have used to bounds some uh
0:09:16	uh experiments
0:09:17	to check this
0:09:18	uh a relationship
0:09:20	so the X axis Q is my man slow which is the uh uh a percentage of a of nonzero
0:09:25	coefficients
0:09:26	and the back Q as this is the number of also
0:09:31	so we see that although the a uh some kind of linear you know relationship
0:09:35	but this is not so strong as we have
0:09:37	it
0:09:38	uh this stinks uh it this is because the
0:09:41	uh
0:09:42	number of so that it's depends not only on the
0:09:46	and number of nonzero coefficients
0:09:48	but also around the distribution
0:09:50	of this coefficient
0:09:53	here in this paper
0:09:54	uh do you want to uh
0:09:57	a find a of which also consider as that distribution of nonzero coefficients
0:10:02	and uh
0:10:03	this that are proposed the with model
0:10:06	for so P information
0:10:07	uh we see here we use the a a a a a number of
0:10:11	uh
0:10:12	this is the number of nonzero macroblock
0:10:15	and that this is number of to macro
0:10:17	and that's a macroblock is
0:10:19	uh a are defined as a macroblock block
0:10:22	you may
0:10:23	or is that i'm as the coefficients at those
0:10:26	so we use
0:10:27	uh this number as an indication of the distribution of the nonzero coefficient
0:10:33	and uh we see on that is
0:10:35	still a constant and the guy i is a parameter we need to update a at that to be during
0:10:40	the encoding
0:10:41	so we seen a uh
0:10:43	i X X is
0:10:45	this is this i time in the back it
0:10:47	and the y-axis is the net but also it's
0:10:50	so we see that uh now there is a strong linear relationship
0:10:54	between this item here and than M L C D
0:10:57	we will use uh this model to estimate the size of the information
0:11:03	so it is uh
0:11:05	is that the two models
0:11:06	uh we propose
0:11:08	and uh now we can introduce our two-stage rate control algorithm
0:11:12	this is that this is basically that's the uh the same
0:11:15	as
0:11:15	uh proposed in the to node or the me but control
0:11:18	model
0:11:19	uh so we have to stay it is
0:11:21	a a first to be a two uh frame level bit allocation so we term how many bits uh i
0:11:26	don't K to to uh in of the frame
0:11:29	and and the analysis that you be do motion estimation of the dct uh and uh transformation
0:11:35	and the be of the coefficients
0:11:37	the that information and the multi information in the back
0:11:40	this is a be later use the by the encoding stage
0:11:43	as the S by the it can draw what
0:11:45	and the in the encoding so is we actually code the or the macroblocks
0:11:49	and the first
0:11:50	uh for each macroblock well we first estimate the had a the number of header bits for the we maybe
0:11:55	macro
0:11:56	except that could be it's for inter macroblock
0:12:00	we could have a clue that the that be be it's because
0:12:03	uh it depends also on the that selected a Q P because if we use different you P the way
0:12:09	that that it's at his number
0:12:11	so uh are that it and number of uh a non-zero coefficients all be different
0:12:15	and the number of to be P is we also be different
0:12:17	so we can only estimate this together is the texture
0:12:21	and the biggest the suspect
0:12:22	uh that is a a a a a a uh a just set aside from that
0:12:26	uh
0:12:27	cut and be bit but it and we get the uh bit about it for
0:12:31	so do be and texture be it's
0:12:32	and the next we find that to the piece to so that's that the sound of their are a number
0:12:38	of so be that be it's
0:12:39	and the the texture bits
0:12:41	uh
0:12:42	that's not exceed as the
0:12:44	come to be the bad
0:12:46	we will use our proposed model to estimate that
0:12:48	and i but also if you bits
0:12:50	and uh uh for text B is to use a a lot of mean with model
0:12:54	and then be used the uh uh uh select a could be you go to each mac rock and after
0:12:59	each map encoding
0:13:01	B B updates uh
0:13:02	prime terms in the risk models
0:13:05	and uh after we have found this for all the macroblocks in the for an is and uh can update
0:13:11	uh a the out but the crime in the model for prediction of the
0:13:15	uh uh next frame
0:13:17	a this is the basic work flow of our proposed
0:13:20	two-stage rate control over them
0:13:22	and we can have a look at some
0:13:25	mm
0:13:26	experiment results
0:13:27	so if you encode the difference because it is that the different uh uh bit rate
0:13:31	and uh here we compare for uh
0:13:34	uh rate control algorithms
0:13:36	that is
0:13:37	and uh rate control next
0:13:40	uh you X two six because our algorithm is implemented in the
0:13:45	uh X
0:13:45	two six four encoder
0:13:47	and this and this by the original automated control
0:13:51	uh without header size estimation
0:13:53	and uh uh for the sake of instead of a and B uh
0:13:58	yeah uh use the wood in the reference paper to estimate the header size
0:14:03	each frame
0:14:05	this is that uh on the second level so first we compare this for uh are within
0:14:10	just four buttons on the second level
0:14:12	is that an sequence that we want to see how close is the
0:14:16	uh actual bit-rate to the high bit-rate
0:14:19	so we see that uh
0:14:21	actually it is uh for rate control algorithms
0:14:25	uh perform well
0:14:26	so we see that the resulting bit rate is very close to the
0:14:30	target bit rate
0:14:31	and uh you insisted we also show some show the psnr
0:14:36	uh compress the
0:14:38	oh that's really gonna to control algorithms with a
0:14:41	there's control E X two things we see that uh
0:14:44	uh for the two
0:14:46	but control
0:14:47	this header size estimation we can achieve
0:14:50	uh
0:14:52	applies to pairs are
0:14:54	this is in the
0:14:55	a second and then we can go down to the
0:14:58	uh for that
0:15:00	uh to see the uh that's speculation of of different uh
0:15:03	and all them
0:15:04	so here we compare
0:15:06	uh
0:15:07	three are the mean that control algorithms
0:15:09	uh
0:15:10	we in of the uh to sequence for about
0:15:13	uh uh and uh the type difference that is
0:15:16	for hundred about
0:15:17	and of see is that of it
0:15:19	just and this is
0:15:20	the uh original are gonna in rate control it out had as that's estimation
0:15:25	and so we see that compare this
0:15:27	just to method of is this does that estimation
0:15:30	uh
0:15:30	the uh ones
0:15:32	you of the frame size here is or
0:15:35	so we see that uh this had is that's estimation you can
0:15:38	uh reduce the
0:15:40	uh for exact calculation bits in the
0:15:42	you
0:15:46	and that we can but the go down to the uh um back
0:15:49	macroblock block that want to see the uh uh Q P variation between a frame
0:15:53	we know
0:15:53	was that a macro can level like control algorithms a lot the two P to be adjusted to
0:15:58	for each mic block so that we can meet the uh type a frame size actually to but if that
0:16:04	you that is changed to match
0:16:06	so we have
0:16:07	um
0:16:08	uh quality activations means in the
0:16:11	for him
0:16:11	so here we simply show some expand the results
0:16:14	uh this adds a fifty cent of had was for him in the for the accused
0:16:20	you can spend that
0:16:21	and uh we compare this really you're the meta control algorithms
0:16:24	the point here is again
0:16:26	for the uh uh
0:16:29	for the original little minute control without
0:16:31	i does that the estimation
0:16:33	so we see that uh
0:16:34	uh uh at the beginning of the frame that you hear meadows
0:16:37	a very large
0:16:38	and that we have and uh
0:16:40	uh for them to be better changes dramatically
0:16:44	we think this is due to the lack of header size estimation
0:16:47	so that at the beginning of a frame
0:16:49	the let control can now to estimate the
0:16:51	resulting
0:16:52	a number of bits actually a T and i the end of the frame H needs to
0:16:56	change this
0:16:57	a dramatically
0:16:59	and we see uh for the two uh our presents
0:17:02	uh this
0:17:03	had a estimation we can achieve a smaller
0:17:06	uh we can you was smaller or of two P relation
0:17:09	for example
0:17:10	this
0:17:11	a our proposed it can draw was them B see that a
0:17:14	uh as to of the whole for the Q P values
0:17:17	you know D that's not change
0:17:19	and he in this for it's changes only be you know very small
0:17:24	so these are a work
0:17:26	i the results
0:17:27	and then now we control
0:17:29	the conclude in
0:17:29	so in this paper we have proposed with the models
0:17:32	for estimation that side of the information you and starts to for
0:17:36	and we also introduce a a two stage me to to control all of them
0:17:40	uh this had a sense estimation
0:17:42	and uh uh had those that's estimation be can achieve a better to control accuracy B can
0:17:47	but you smell of frames that's fluctuation tuition in the
0:17:50	sequence
0:17:51	and the can can also achieve
0:17:52	smaller or to be variation within that
0:17:56	okay i think this
0:17:57	or or met up to
0:18:01	we we can do have a couple questions
0:18:08	so you compute
0:18:11	yeah
0:18:12	yeah you compared
0:18:14	i
0:18:15	well
0:18:17	i
0:18:18	yeah
0:18:20	yeah
0:18:21	yeah so so
0:18:22	this one
0:18:24	the the the
0:18:25	the right car is for the
0:18:28	this this
0:18:28	the
0:18:30	yeah
0:18:31	but in the weapon paper actually the for the source it's
0:18:34	use the
0:18:36	uh the quadratic model
0:18:38	not a lot of model
0:18:40	yeah we have used them and this model
0:18:42	the together with
0:18:43	not only mode
0:18:45	this is
0:18:46	experiment results
0:18:48	does
0:18:51	it
0:18:52	also works
0:18:54	also works
0:19:07	so it was that you straight
0:19:09	the extreme
0:19:12	the that's you extract you
0:19:14	the uh okay so i in this we encode the a uh for for each seconds to being called the
0:19:18	for to uh uh uh a three hundred or a frame
0:19:21	but only the fourth
0:19:22	that's that's uh i four
0:19:24	for the following a
0:19:26	a a a whole with different
0:19:28	oh
0:19:29	be
0:19:29	hmmm
0:19:30	oh i hope will be you
0:19:33	uh uh
0:19:35	uh we we do not use any before
0:19:38	so we do not use and hierarchical
0:19:41	oh
0:19:59	so
0:19:59	uh_huh
0:20:01	you
0:20:01	usually larger size
0:20:04	uh
0:20:05	however
0:20:05	yeah
0:20:06	uh the only tried uh cues if the consistency
0:20:11	we get a result
0:20:14	what
0:20:17	still
0:20:19	hi
0:20:19	which
0:20:20	uh
0:20:21	if
0:20:22	okay because
0:20:23	also do that
0:20:26	i think for a uh
0:20:28	for higher resolution depends on the target for a target a bit rate
0:20:34	so
0:20:36	maybe if the if the uh target uh
0:20:39	target bit rate is
0:20:40	i
0:20:41	i think
0:20:42	uh picture piecemeal
0:20:44	okay cap most of the
0:20:46	basing the sequence
0:20:47	so
0:20:48	so how does that
0:20:49	mission
0:20:50	will be nice
0:20:52	efficient
0:20:52	but the of the
0:20:54	uh
0:20:54	bit-rate this but
0:20:56	we don't in the middle
0:20:57	little
0:20:59	i think that uh
0:21:00	uh
0:21:01	how does that
0:21:01	estimation
0:21:02	we bring
0:21:03	a lot of
0:21:04	okay
0:21:06	yeah
0:21:08	you
0:21:09	i
0:21:11	if
0:21:12	i
0:21:14	as you
0:21:15	which
0:21:18	i
0:21:20	you know
0:21:21	yeah
0:21:23	i
0:21:25	oh
0:21:27	actually a we we haven't a need uh a high resolution
0:21:31	yeah
0:21:32	you haven't done on a a i am and the for high resolution sequence
0:21:35	but i think uh
0:21:37	for
0:21:37	is
0:21:40	the the percentage of nonzero coefficients is also
0:21:45	so uh a a one hundred depends on the side of the frame
0:21:48	on the other hand it's
0:21:49	it's also depend on the target before
0:21:52	so you you want to do
0:21:54	right
0:21:54	the high uh
0:21:55	hi uh resolution second
0:21:57	with a you are there is slow
0:22:00	a only a little bit rate
0:22:03	you you can to you
0:22:05	i speech
0:22:06	uh_huh you
0:22:08	it
0:22:09	yeah just
0:22:10	it makes it
0:22:12	yeah sure
0:22:13	so
0:22:13	this this case
0:22:14	i think that uh
0:22:15	um
0:22:16	maybe the
0:22:17	percentage of nonzero
0:22:19	you you higher are much higher
0:22:21	then our experiment
0:22:24	okay okay

IMPROVED P-DOMAIN RATE CONTROL WITH ACCURATE HEADER SIZE ESTIMATION

Video Coding

Přednášející: Fan Zhang, Autoři: Fan Zhang, Eckehard Steinbach, Technische Universitaet Muenchen, Germany