Přepis řeči - EXTENDED INTER-VIEW DIRECT MODE FOR MULTIVIEW VIDEO CODING

0:00:13	you
0:00:16	one
0:00:18	exactly
0:00:18	and a boring points in tandem one
0:00:20	my name's got the carnage the
0:00:22	uh uh as it was set i from one diversity of technology
0:00:26	and it's a cloud that rate rider to a what can not today
0:00:30	and
0:00:31	in my
0:00:31	and i would discuss
0:00:33	the
0:00:34	extended inter view direct mode for the multi view video coding
0:00:39	so to start with
0:00:40	uh
0:00:41	development of uh efficient it's a solutions for multi view video coding becomes
0:00:46	more and more
0:00:48	a an important these days
0:00:49	as the the D you video applications are getting
0:00:54	and more and more popular
0:00:56	and in the standard to a solution no as
0:01:00	the and X H of the a C H two
0:01:03	six six four standard
0:01:05	and a simple but the quite efficient a my it
0:01:09	is you
0:01:10	and
0:01:11	in this
0:01:12	standard to the solution
0:01:14	and a not in order not to a a sense uh
0:01:18	a a independent you
0:01:20	so
0:01:21	in order not to and got independent
0:01:24	and every you of the multi you sequence
0:01:26	and we uh uh i i a a reference pictures from other uh you
0:01:32	to that
0:01:33	uh reckon
0:01:34	which are list
0:01:35	so
0:01:36	this is the main uh
0:01:37	the domain
0:01:38	and in there view my congressman man
0:01:41	which is used in the multi
0:01:42	you got me
0:01:44	and unfortunately there
0:01:46	read a then C be between a a a bit streams of the neighbouring use
0:01:51	a is still quite a big so a uh
0:01:54	phil four
0:01:55	further problem exist
0:01:57	and on the other hand
0:02:00	and
0:02:01	there is
0:02:02	some some additional information the multiview view video which can be adopted
0:02:07	a a to uh decrease this redundancy between a a bit streams in
0:02:12	they bring use
0:02:13	and this information is
0:02:15	that that information
0:02:18	that that well for the which are described where that three D geometry of the scene
0:02:23	and a if we at all
0:02:25	this is information in encoding process
0:02:28	and we can get that uh
0:02:31	that
0:02:32	performance of the multiview view video codec
0:02:35	and and should be
0:02:36	bigger or and as a result
0:02:39	the result and the
0:02:40	bit strange should reduce
0:02:43	so the main idea
0:02:45	presentation
0:02:46	still
0:02:46	use that
0:02:48	that information to improve the compression of the multiview video
0:02:53	and
0:02:53	use the use that's of that that information beeps i'll some uh a new possibilities
0:03:00	a for prediction
0:03:02	as we all know that you all the bit stream
0:03:05	contains
0:03:06	most the of the control data prediction error or
0:03:09	motion data
0:03:10	reducing using a a you two
0:03:12	prediction
0:03:13	and a
0:03:15	and he of this uh uh that's of data thus
0:03:17	it's the
0:03:19	as uh
0:03:20	make a a bit stream
0:03:21	smaller smaller
0:03:22	a a two day i'll discuss only prediction of the motion data
0:03:27	to decrease their
0:03:28	bit right
0:03:30	oh
0:03:31	what to let let's imagine that to
0:03:34	we've got to be used to and go
0:03:36	the reference view which was already in that the bit stream
0:03:40	all this
0:03:41	you was
0:03:41	know
0:03:43	at the time of uh a ink the other view
0:03:47	and to
0:03:49	what we are uh actually try to do is to predict the motion
0:03:53	a data
0:03:54	from the neighbouring you
0:03:56	and using
0:03:57	the three D dependencies between objects
0:04:00	a C
0:04:01	for T
0:04:01	a dependencies are described by a that information
0:04:07	oh yeah
0:04:08	a bit more details
0:04:10	and
0:04:12	first of all of like to point out that
0:04:14	the prediction of motion i i data from the right
0:04:17	you
0:04:18	the the independently for each point of and to
0:04:22	and
0:04:23	of and code that be a picture
0:04:26	and
0:04:28	we we are to only using that the that information of the ref
0:04:32	you only
0:04:34	a we check that that volume for that
0:04:36	for each of the that
0:04:38	a points in the river
0:04:40	you
0:04:40	and based on this information
0:04:43	and
0:04:44	we project
0:04:45	to a a location in that feed this space using that that image based rendering
0:04:51	so we get a
0:04:52	a point location of pretty space
0:04:55	and next
0:04:56	we re project is position in
0:04:59	and
0:05:00	but that's to picture
0:05:02	and
0:05:03	and
0:05:04	as a result we get to i
0:05:06	a of a point
0:05:07	one in the right
0:05:08	you and one call it's you'll
0:05:10	and this uh points are a part
0:05:14	or so
0:05:15	some uh a connection between that
0:05:18	exist
0:05:19	uh
0:05:20	and it's it as there
0:05:22	a a
0:05:23	motion data for the pointing direction
0:05:26	you is already known because
0:05:28	you was already in bit
0:05:30	so we can check the motion vectors selector
0:05:32	speech or in
0:05:33	all
0:05:34	a which are
0:05:35	uh in that block
0:05:37	uh
0:05:38	for this
0:05:38	or
0:05:39	and uh
0:05:41	simply the right that two
0:05:43	and describe
0:05:44	the motion vectors and reference picture dies
0:05:47	of that point
0:05:48	in the column view
0:05:49	so this
0:05:50	a the main idea of the
0:05:52	of
0:05:53	a problem
0:05:54	and evaluate
0:05:56	uh this idea
0:05:58	and
0:05:59	we integrate it's with that and B C motion video coding
0:06:03	and a reference scope there
0:06:06	and and the plantation was
0:06:08	the
0:06:08	uh uh
0:06:09	we do that
0:06:11	uh S and you might have model
0:06:13	a video compression
0:06:15	and
0:06:16	which we call the extended inter view direct mode
0:06:19	so yeah i D
0:06:21	uh
0:06:23	now it's uh look how a dog to the a and B C a coding scheme
0:06:29	and that's a that we've got the preview
0:06:32	uh uh two and go
0:06:33	that
0:06:34	basic scheme which is used in and
0:06:36	see
0:06:37	it's more or less like
0:06:38	they want to picture
0:06:40	so we've got a a a a base you
0:06:42	this C zero which is
0:06:44	and and is the first well so for that based you there are null
0:06:48	error
0:06:49	you
0:06:49	uh a available
0:06:51	and there are two a more you
0:06:54	see C two and you see well
0:06:57	a i each of
0:06:58	uh which are and that with
0:07:00	one or two ref
0:07:02	you
0:07:03	i would like to
0:07:04	a
0:07:05	point out that it's uh
0:07:07	for that use C tool the inter view a reference
0:07:11	the
0:07:12	uh which is
0:07:13	the used due to this scheme
0:07:15	is only for that
0:07:17	and it for the the on or picture areas
0:07:20	so for that
0:07:21	time instance T zero
0:07:23	uh place look at the picture
0:07:25	for uh a time instance
0:07:26	do you want to and others
0:07:29	there is no our between you use C zero and C two
0:07:34	and
0:07:36	in contrast to for you C while
0:07:38	a every a picture
0:07:42	for a every
0:07:42	time
0:07:43	as there are
0:07:44	are are between use C zero and to C one or
0:07:48	you C two and C one
0:07:51	uh
0:07:52	however
0:07:53	uh the only uh
0:07:55	the only uh what if a question do the income comparison with
0:08:00	stand
0:08:01	a simple cost and called is that there right
0:08:04	at least
0:08:04	the the are modified
0:08:07	in the way that it for each time in
0:08:10	star
0:08:11	uh
0:08:12	features from a from
0:08:13	use from thing but
0:08:14	a from the neighbouring used are at their uh to that
0:08:18	a rip
0:08:19	which lee
0:08:21	yeah
0:08:22	after i think uh there
0:08:24	i i E E I D mode two
0:08:27	and and this
0:08:28	the scheme
0:08:30	we get some you uh
0:08:32	prediction possibilities
0:08:34	and i i would like to point out that
0:08:36	this uh
0:08:37	a are a smart that's right
0:08:40	i i do not to a modified the referee
0:08:43	features list
0:08:45	a a as we don't the use
0:08:47	a a new pictures
0:08:49	and
0:08:50	this
0:08:51	this are was only a
0:08:53	tell that
0:08:55	a
0:08:55	for this features we can is
0:08:59	we can predict the motion information from other views using to good
0:09:05	and
0:09:06	what this where not not no
0:09:08	also is that
0:09:09	then you model
0:09:11	used to only for that
0:09:13	not not you
0:09:15	because a based you
0:09:16	no right
0:09:17	you
0:09:19	so there is no you to two
0:09:21	to to uh to refer
0:09:24	and the other of thing is that i
0:09:27	new model can be applied only for that
0:09:30	and on on for pictures because in none or pictures
0:09:33	and there
0:09:34	motion dot doubles
0:09:36	we can read
0:09:36	do
0:09:37	and
0:09:38	it's only that
0:09:40	uh
0:09:41	the the most but those uh describe that
0:09:44	and uh
0:09:45	them that the motion which is not collected
0:09:48	a a different time instants
0:09:50	well only two
0:09:51	and different C you
0:09:53	so
0:09:54	and so we don't to use that
0:09:57	and
0:09:58	now a few words about the evaluation of the i
0:10:01	the model
0:10:03	and and we but
0:10:04	de the I V D model
0:10:06	with if too codecs
0:10:08	and this is
0:10:09	then B C four point all which is the reference
0:10:12	the
0:10:13	reference software
0:10:15	and
0:10:16	the previous one of the previous version
0:10:18	of
0:10:19	of
0:10:19	a a of the software
0:10:21	which also
0:10:22	okay case the motion skip tool
0:10:24	a adopted in that
0:10:26	most speed
0:10:27	a according to which
0:10:29	use this uh is a similar idea all
0:10:31	a a a pretty thing motion information from a neighbouring bureaus
0:10:35	a cover it does not to the use
0:10:38	and information about that or three D G
0:10:42	so it's uh and an interest
0:10:44	a them to this
0:10:46	and
0:10:47	in our test we used
0:10:49	five different this sequences
0:10:50	who book i've by parallel where
0:10:53	as something power and of mine
0:10:56	and
0:10:57	the results were obtained for to be value
0:11:00	for four for the six and four do
0:11:03	and the last thing is that
0:11:05	bit stream for for uh
0:11:08	which are present
0:11:10	uh are the bit streams
0:11:11	the the for a single you so we only
0:11:14	a
0:11:15	so i would show only a results for the call you would vol
0:11:20	and we've out the reference you
0:11:22	bit stream or we about the
0:11:25	bit streams to encode the data
0:11:28	because in our approach we assume that
0:11:30	and that that uh information already
0:11:33	and i by to what called or
0:11:35	for some other purpose
0:11:38	so we don't do not include it's uh and uh uh uh the the B
0:11:43	all the results
0:11:46	and
0:11:47	because uh in the reference uh
0:11:50	use the sorry
0:11:51	the ref
0:11:52	multi well all this not this material
0:11:55	there were only and three views available
0:11:58	so there are a two possible scenarios to to check
0:12:02	first number is
0:12:03	that we are coding to you C two
0:12:06	and and the reference you is to the base use of the use C zero
0:12:12	and the other scenario is
0:12:13	the encoding coding all use you one
0:12:16	and a of the reference you C zero
0:12:21	uh
0:12:21	and finally the results
0:12:23	so
0:12:24	as you can see and you do you
0:12:27	yeah i don't think the i i yeah i i D mode into D
0:12:30	a a a a and and P C software
0:12:33	and
0:12:34	get
0:12:34	some uh bit-rate rate reduction
0:12:38	a in in there's the
0:12:40	there's compare them with uh the referee
0:12:43	we we you reference software
0:12:45	and uh as we can see
0:12:47	that
0:12:47	be to bit reduction is
0:12:50	uh for every at test sequence
0:12:53	uh
0:12:54	the are range one is
0:12:55	that are part of the ref
0:12:56	software
0:12:57	with the motion speed or a label
0:13:01	uh
0:13:02	we also not is
0:13:03	that
0:13:03	the a a bit reduction
0:13:06	and case of snow or two
0:13:08	and
0:13:09	compared of the reference software
0:13:11	uh
0:13:13	blue one
0:13:14	uh
0:13:15	also gives us a a a a a bit rate
0:13:18	bit-rate rate reduction
0:13:20	a cow what are in case of the
0:13:22	a compress on the motion he
0:13:24	we see that a there is a
0:13:27	uh as sequence
0:13:28	a at i'm maturity
0:13:30	a a for which a the motion keep for the
0:13:33	what but they're
0:13:34	yes
0:13:35	uh
0:13:36	to summarise the result
0:13:38	uh
0:13:39	as we notice the extent
0:13:41	a inter view direct mode improves compression surely
0:13:45	and
0:13:46	against reference software
0:13:48	we get
0:13:49	this
0:13:49	bit-rate reduction for all this see test signal
0:13:52	for or bit rate
0:13:54	and
0:13:55	and the average bit-rate savings were uh
0:13:58	six six point nine percent
0:14:00	a now to and
0:14:01	point five point
0:14:03	three percent
0:14:04	in case of standard two
0:14:06	and this
0:14:07	you vol are have weighted
0:14:09	to
0:14:11	based
0:14:11	the instrument we yes
0:14:14	uh
0:14:15	and compares them with the motion skip tool
0:14:18	i we get the bit rate reduction from all this
0:14:22	were checked
0:14:23	and uh and the average bit which
0:14:25	uh is savings
0:14:26	are eleven point three percent
0:14:28	and
0:14:29	two point five percent
0:14:31	for never two and
0:14:33	a so i one and i two
0:14:35	uh
0:14:36	lean
0:14:38	to conclude my
0:14:39	and
0:14:40	and a
0:14:42	we shall that that information
0:14:44	can be efficiently use
0:14:46	to represent motion data
0:14:49	and in multi view video
0:14:50	a bit streams
0:14:53	and
0:14:53	do do you use that of the
0:14:55	three young geometry we get a an Q right
0:14:58	the of motion data from the you
0:15:02	uh
0:15:03	the whole algorithm
0:15:04	a interest
0:15:06	the the computation going with
0:15:07	complexity of be called in english lee
0:15:11	as so the this is true but in
0:15:13	and and so the the plantation
0:15:16	oh
0:15:17	the
0:15:17	new idea was
0:15:19	and the for the extended inter view direct mode
0:15:22	uh
0:15:23	as nice
0:15:24	as we so the compression that used to see
0:15:26	a improvement was
0:15:28	uh a two point blank
0:15:30	to let and to three percent
0:15:32	uh
0:15:33	we call it can
0:15:34	we draw is uh the more accurate that model will uh be a
0:15:40	and the last but not the least
0:15:42	the use it
0:15:43	oh that
0:15:44	from
0:15:44	a such associated with
0:15:46	uh
0:15:47	and uh the reference you only
0:15:50	and makes this solution adaptable able to than most of the multi view video applications so
0:15:55	applications where for example texture and that are and called
0:15:59	the
0:16:00	to together
0:16:01	dependent the
0:16:02	and
0:16:03	and
0:16:04	this
0:16:04	this this solution is also applied to
0:16:08	no
0:16:09	so thank you very much
0:16:11	real a
0:16:18	you have questions
0:16:24	so
0:16:25	like
0:16:27	what
0:16:30	me
0:16:33	what sounds like
0:16:36	you
0:16:38	and the reason we we should to the we should note that the
0:16:43	it increased
0:16:44	uh
0:16:45	the
0:16:45	complex you of the you gotta because
0:16:48	some additional orders
0:16:50	the is you'll still a to project the pixel location
0:16:54	and and some but uh
0:16:56	some calculations are uh that but
0:16:59	this is not much
0:17:01	so
0:17:01	what
0:17:04	if you
0:17:05	i
0:17:06	and and are not not not this some that
0:17:12	any questions
0:17:14	i
0:17:16	i think you're

EXTENDED INTER-VIEW DIRECT MODE FOR MULTIVIEW VIDEO CODING

Stereoscopic and 3-D Coding

Přednášející: Jacek Konieczny, Autoři: Jacek Konieczny, Marek Domanski, Poznan University of Technology, Poland