Speech Transcript - OPTIMIZED EDGE APPEARANCE PROBABILITY FOR COOPERATIVE LOCALIZATION BASED ON TREE-REWEIGHTED NONPARAMETRIC BELIEF PROPAGATION

0:00:14	a low uh i'm going to
0:00:16	but then you
0:00:18	uh
0:00:19	our
0:00:20	the localisation
0:00:21	based is not that are really a where we mainly focus on
0:00:26	but an empirical facts for optimal
0:00:28	and to a really
0:00:30	are
0:00:31	the the
0:00:32	uh
0:00:33	yeah for part right is the from times
0:00:36	no
0:00:36	the right panel
0:00:38	then
0:00:38	three and uh and of that
0:00:40	probably most of the my
0:00:45	that
0:00:46	with the soft introduction
0:00:48	uh
0:00:49	i i don't model for
0:00:51	oh
0:00:52	that is uh
0:00:54	introduced to a it not by really for is uh
0:00:57	and then main part of a
0:01:00	it is
0:01:01	really
0:01:01	as we do the using monte carlo simulation
0:01:04	and
0:01:05	and
0:01:05	loose
0:01:09	so a some real problem do not not farm the hz now
0:01:13	but but is that for
0:01:15	a message passing method for vol
0:01:18	for a number of the case of to think of
0:01:20	there is a
0:01:22	a a this is a a a a a a bit is
0:01:24	based on that are used
0:01:27	i no
0:01:28	uh can don't know number of the condition
0:01:30	like a piece without didn't
0:01:32	uh
0:01:32	the the the yeah
0:01:34	and many more
0:01:38	however the main problem of this method is
0:01:41	uh
0:01:42	or or or or that really
0:01:44	this means that there are quite a them is really that
0:01:47	that
0:01:48	so
0:01:49	and also possible not or to score
0:01:51	and
0:01:52	not covered
0:01:53	you know
0:01:54	not problem for of like this at least we have not
0:01:57	this brown
0:01:59	uh
0:02:01	there are a lot of solutions
0:02:02	course
0:02:03	a standard was that a to partition
0:02:06	uh
0:02:07	is also possible to sort by splitting the finding season
0:02:10	and
0:02:11	uh
0:02:13	so our research and or
0:02:15	to our work
0:02:17	some kind of a
0:02:18	oh submission now
0:02:20	mean
0:02:21	simplest method
0:02:23	so
0:02:24	that are is to it
0:02:26	yeah
0:02:27	iteration
0:02:27	not part of
0:02:28	are
0:02:28	however
0:02:30	if does not very
0:02:31	serious problem because
0:02:33	oh or anything
0:02:34	and it is
0:02:37	the the parameter
0:02:39	sure
0:02:39	or as the probability that a
0:02:41	and
0:02:42	years
0:02:43	and and with a three
0:02:46	yeah be a best
0:02:48	oh
0:02:48	so
0:02:50	right
0:02:54	and so we propose a one of of
0:02:57	this method for
0:02:58	as it is
0:02:59	same
0:02:59	and
0:03:00	really
0:03:01	for
0:03:03	we
0:03:04	find a a possible
0:03:07	and minimize the error
0:03:09	mean
0:03:09	to really an estimate
0:03:13	so
0:03:15	i real star
0:03:16	uh
0:03:16	simple
0:03:18	a a a a a little additional
0:03:20	uh uh
0:03:22	yeah you or or or time you know
0:03:25	lee
0:03:26	yeah
0:03:27	the the a distance measure
0:03:29	we you know probably or not not well
0:03:31	all
0:03:33	uh this is for this problem so that the and it's no
0:03:37	yeah it's on
0:03:38	so
0:03:40	a a you know the two
0:03:42	oh
0:03:43	and and they are of this these they need to
0:03:46	um are at each other but are not that money
0:03:50	are or not
0:03:51	and
0:03:52	except increased
0:03:53	speech and i of this we make this method that
0:03:59	or got this this person you not but those are some known
0:04:03	and the
0:04:04	uh
0:04:07	the noise be an four and five years used
0:04:10	or
0:04:11	we
0:04:11	uh a you know the a samples of then
0:04:14	yeah
0:04:15	there's not sound
0:04:16	no for this problem we can use was
0:04:18	we don't want
0:04:19	or and uh
0:04:21	the problem with north wind
0:04:24	for these problem
0:04:25	recall this
0:04:28	um and you also need to find probability of detection probability is the no D
0:04:33	when you you know you
0:04:36	this is a i don't not know the one
0:04:39	no no you
0:04:40	only you
0:04:41	uh there there is the
0:04:43	the distance is smaller than speech right a
0:04:47	so and have a model
0:04:49	a a right form a very
0:04:52	uh
0:04:53	i defined also indicate to the action of by it
0:04:57	right a
0:04:59	there is additional but
0:05:01	yeah can like
0:05:02	i
0:05:02	yeah
0:05:04	uh
0:05:05	a of all the right or and all the light
0:05:13	or are we are going to different problem in
0:05:15	and that all models
0:05:16	we are going to use
0:05:17	mark that one hundred know one
0:05:20	just more suitable for this problem
0:05:22	uh that means that the
0:05:24	uh B
0:05:25	for it's no design right prior
0:05:29	a not that things you know
0:05:31	i are from that one
0:05:33	i i i don't is but then she of three
0:05:35	like
0:05:36	a function of one mission
0:05:39	so
0:05:41	and define this
0:05:42	we can see that for a attention there everything things that and want to know and
0:05:46	i the best one oh
0:05:48	the prior
0:05:49	is
0:05:50	still
0:05:51	so
0:05:53	for we can i
0:05:54	the posterior here
0:05:56	right
0:05:56	everything of all single not your
0:05:59	there is potentially a
0:06:01	so we can going to push a button or well that's fine that your money you know
0:06:06	uh
0:06:08	or as it is
0:06:10	oh the position and
0:06:11	so this is usually that it was a really use less passing met
0:06:16	where there
0:06:17	and uh also a mission
0:06:20	then show that they have these time so i think that that's
0:06:23	writer
0:06:24	and here is but is so is the line
0:06:26	function
0:06:28	this probably in his probably due to case
0:06:31	the
0:06:32	a probability of the detection and
0:06:34	a likely with the position of the northeast data
0:06:37	you
0:06:38	yes i don't no
0:06:39	this is not a fixed and just might be but
0:06:43	and you are not going to use a a is this you like this only the graph
0:06:48	and uh
0:06:49	uh
0:06:50	this this right and
0:06:51	means means that we need more
0:06:53	communication
0:06:54	not
0:06:55	smart minute
0:06:56	range
0:06:57	so
0:07:00	i mentions so that they is that in this addition the joint for really this is not a
0:07:05	uh that that was so
0:07:06	we need to use message passing that that
0:07:08	and
0:07:09	the main problem
0:07:10	it is
0:07:12	uh
0:07:13	but this is not a scene
0:07:15	little loss
0:07:16	so it that if you can actually you can have like that but
0:07:20	you know yeah that are there is no
0:07:24	i don't see that it
0:07:25	cover and that
0:07:26	provide
0:07:27	a X can also
0:07:29	do you are not
0:07:31	the
0:07:32	and
0:07:34	use right of to your rate partition
0:07:37	a given by uh
0:07:39	really
0:07:40	a and
0:07:42	a a message data
0:07:44	first i
0:07:45	i i start at the location of a special case of given to this and to be they are really
0:07:50	here
0:07:52	you to
0:07:53	well
0:07:53	so that i
0:07:56	really locations so
0:07:58	was that you're but you know it i we you really is
0:08:01	right of all the right or on incoming that
0:08:06	and the
0:08:08	message from node you know T is the
0:08:12	we found out
0:08:14	you about position of T
0:08:16	so
0:08:17	i want you know
0:08:18	so really
0:08:19	know
0:08:20	information about these that
0:08:22	so we make this problem may right
0:08:26	to know read out
0:08:28	no
0:08:28	the value that you don't know so
0:08:31	you only information about the about
0:08:34	engine out
0:08:36	and you also you need to divide
0:08:38	uh
0:08:41	really
0:08:41	a uh with the
0:08:43	be having events
0:08:44	we use the variation because
0:08:46	uh the
0:08:48	you need to double count the election
0:08:51	we we direction
0:08:52	used
0:08:53	is
0:08:53	one
0:08:57	oh what does not lot to lose this is not a wiener problem so
0:09:00	and this is also a
0:09:01	you are a um very big that's so we need to use
0:09:04	uh uh part
0:09:06	dimensions
0:09:07	are we don't know what that the you can take the reference this this that the same as not are
0:09:12	really relation
0:09:14	so are final really
0:09:17	are a set of five
0:09:19	uh
0:09:22	a part of that and all that these are all
0:09:24	is that right
0:09:26	that i mean
0:09:27	as we doesn't right race
0:09:29	i this power so
0:09:30	there's is no problem with this
0:09:34	uh kind of the main vector
0:09:36	are that are to the
0:09:37	and you didn't
0:09:39	really do
0:09:40	is the probability that a a it is not a
0:09:42	i'm straight
0:09:44	uh for example
0:09:45	that that is do you have just last time three so
0:09:48	uh well
0:09:49	equal to one
0:09:52	and that's another stop
0:09:54	is
0:09:55	for no please
0:09:56	so here we can sixteen and say
0:09:59	so that is
0:10:00	yeah about it
0:10:01	you seen eight of them
0:10:02	so
0:10:03	um
0:10:05	the
0:10:06	and you know for the zero or is a equal to zero point five
0:10:11	you do not assume at this
0:10:13	a for all that
0:10:15	and
0:10:17	and so
0:10:21	as you can see that a lot of spongy you for small that so it is not so
0:10:26	uh
0:10:27	possible was able to find a for each edge
0:10:29	that this
0:10:31	yeah and probability
0:10:32	set
0:10:33	it not seen cases
0:10:34	so you are going to
0:10:37	remove this restriction and to my is uh it is you
0:10:41	or not
0:10:42	like there's and okay
0:10:43	like
0:10:44	but
0:10:45	and the location is also used in values and
0:10:49	or are all equal to one
0:10:50	so
0:10:51	it
0:10:52	instead of the here by two
0:10:54	lower an optimal all minimizing kl divergence
0:10:57	with
0:10:59	to really find a a set and the
0:11:02	the
0:11:03	also a are was there a what of the problem is that
0:11:06	uh finding can be really and the
0:11:10	to do it
0:11:12	not i
0:11:13	but can simply is
0:11:14	in
0:11:16	so we we want to use
0:11:18	R S I error
0:11:19	but means that all
0:11:21	uh
0:11:25	or you start for not only
0:11:27	but in one thing you one
0:11:29	find him that
0:11:32	is that it
0:11:33	okay a six it is one dimensional case
0:11:35	and if you have a lot of but is not relevant for this
0:11:40	problem
0:11:41	so we but here
0:11:43	yeah were there
0:11:45	and it is probability
0:11:47	and
0:11:48	mean square error line
0:11:50	so we can see that
0:11:52	yeah
0:11:53	optimal it
0:11:54	or or a zero point five
0:11:56	and also
0:11:57	saying that the remaining lines
0:11:59	root square means or or or or or uh
0:12:02	yeah yeah
0:12:04	a
0:12:05	you you part of getting what is special case
0:12:08	and it is that a to i
0:12:10	you can see
0:12:11	okay
0:12:12	yeah can the egg
0:12:14	okay they were does not
0:12:16	very
0:12:17	uh
0:12:18	see again
0:12:19	from
0:12:20	root mean and uh uh
0:12:22	uh
0:12:23	so maybe this the main problem by people use your to do not part
0:12:27	really evaluation for any
0:12:29	there is usually the only take up one point estimate you
0:12:32	yeah a very similar problem
0:12:34	but
0:12:34	the leave will be over all and you can see this
0:12:38	a in this example here we can see a go was you know and
0:12:41	at really more
0:12:43	uh
0:12:46	i really don't have a really but it's really and
0:12:50	and
0:12:51	to read and
0:12:52	the the location of that as you can see
0:12:54	uh and it really
0:12:57	is
0:12:57	very you know that one
0:12:58	it's very tight
0:13:02	not call a presentation of or is it
0:13:05	and to the right don't time really propagation
0:13:08	something between
0:13:10	a and the
0:13:11	uh
0:13:12	then be be really so it is that
0:13:15	it's not the best
0:13:15	because it one is that
0:13:17	estimate
0:13:18	we you need to a apply for example do to be read
0:13:21	then
0:13:23	because you that we can never make
0:13:25	yeah
0:13:26	i zero
0:13:28	using this approach
0:13:31	so that would use you is that we can minimize
0:13:34	yeah that the
0:13:36	yeah they were
0:13:38	as you need to begin yeah because
0:13:40	we're is you a so if you have that one believes that
0:13:44	uh
0:13:45	make
0:13:46	from
0:13:47	so are we
0:13:48	move to
0:13:49	then to find a that for
0:13:50	in the politics
0:13:52	uh
0:13:54	the same or
0:13:55	you you know noise
0:13:58	oh
0:13:59	a not anybody know
0:14:01	but
0:14:02	i
0:14:03	this
0:14:03	a is not but then
0:14:05	so not a lot
0:14:06	we are using vectors
0:14:08	okay so i
0:14:09	of course
0:14:10	and i the
0:14:13	for them very problem for each
0:14:15	not legally
0:14:17	so
0:14:18	a lot
0:14:19	oh the role of that part of that one of these
0:14:22	is
0:14:23	a national
0:14:25	so
0:14:27	yeah
0:14:29	this all
0:14:29	S or at least you one or at
0:14:32	but
0:14:33	this is on a segment code
0:14:36	so you also sort of think that or
0:14:39	uh
0:14:43	the that
0:14:44	everything
0:14:44	is that
0:14:45	okay
0:14:45	you it's life
0:14:46	not difficult but i don't where
0:14:49	so but um it's
0:14:51	but uh
0:14:53	oh
0:14:54	similar or of the like that that
0:14:56	that different
0:14:58	firemen
0:14:59	these
0:15:00	i
0:15:03	so that you can say that a a a a a lot of wants to be the fence on are
0:15:08	originally
0:15:09	not on some either
0:15:11	that
0:15:12	the net
0:15:13	is
0:15:14	and not
0:15:14	we didn't so everything
0:15:17	paper
0:15:20	with the but was also a little
0:15:23	a little bit
0:15:24	uh so we propose by
0:15:26	so
0:15:28	uh not part
0:15:30	partition
0:15:30	and i four for i of a show
0:15:34	oh
0:15:35	one also has it's
0:15:37	very it performance
0:15:39	because for a
0:15:40	so i
0:15:41	error that
0:15:42	yeah the for example
0:15:44	case
0:15:46	uh
0:15:49	on the the to model depends on you was not it's not really
0:15:54	uh
0:15:55	a a net
0:15:56	duration does not have
0:15:58	but on the network
0:16:01	and the
0:16:03	that
0:16:04	can find a document
0:16:06	and can these two area
0:16:09	uh
0:16:10	that's
0:16:11	is is as the same
0:16:12	oh this method of does need to write
0:16:14	right
0:16:14	all
0:16:15	well
0:16:17	and also we need to
0:16:18	i i it's no that
0:16:21	i this using it
0:16:22	send
0:16:23	a
0:16:24	for example but
0:16:25	what to know that we cannot use lots or not
0:16:28	yeah
0:16:29	start
0:16:30	is
0:16:30	you have a lot of neighbours and all
0:16:32	that's that's i mean
0:16:34	uh i one
0:16:35	i can make a wrong decision that
0:16:37	a wrong not as well but
0:16:38	is that although
0:16:40	i are also so
0:16:42	a
0:16:43	is a lot of
0:16:44	note
0:16:46	a of the net
0:16:50	and
0:16:50	seems to you sort this paper
0:16:52	for
0:16:54	that's to mention
0:16:55	it
0:16:55	a a similar form
0:16:57	a a of fine for all point be possible as something
0:17:00	it's
0:17:01	one can
0:17:02	is
0:17:03	or
0:17:04	additionally as the little so
0:17:06	report
0:17:08	two button
0:17:09	role in a low dimensional space and then for like in
0:17:12	i
0:17:17	so are also i that also mention and
0:17:20	that's not is that
0:17:22	and not like a shot
0:17:24	first one
0:17:25	this is
0:17:26	is this
0:17:27	it is that a
0:17:28	my
0:17:29	really
0:17:29	okay
0:17:30	you
0:17:31	i
0:17:31	is it a little
0:17:33	at this
0:17:34	and it really
0:17:35	and also what
0:17:38	uh
0:17:38	gain to the standard net
0:17:40	and say that the product
0:17:42	we apply applied
0:17:43	the the this of for ldpc codes
0:17:46	this case
0:17:47	yeah
0:17:47	i i the next
0:17:49	also be
0:17:51	not in for is beyond that
0:17:53	oh
0:17:54	yeah and so we did
0:17:55	the the
0:17:57	oh
0:17:59	you know case
0:17:59	three
0:18:00	oh but four a B B or
0:18:02	and and be like in case
0:18:04	is there is
0:18:07	so that you know thing
0:18:08	i
0:18:22	uh here
0:18:23	i
0:18:24	do could be this
0:18:26	and uh it but you know that it i
0:18:28	i
0:18:29	yeah but
0:18:30	additionally expense
0:18:44	a to the resin scaling so
0:18:48	this would
0:18:49	but here we have a
0:18:51	this is completely different for you get this output
0:18:54	a probability density function is probably i guess
0:18:57	only you
0:18:58	uh
0:18:59	estimate of the with this
0:19:01	so
0:19:02	it could be optimal Q the gaussian noise
0:19:05	but here we can be completely independent can i
0:19:09	and you noise
0:19:11	you
0:19:13	this is a lot more expensive
0:19:15	um
0:19:16	the

OPTIMIZED EDGE APPEARANCE PROBABILITY FOR COOPERATIVE LOCALIZATION BASED ON TREE-REWEIGHTED NONPARAMETRIC BELIEF PROPAGATION

Distributed and Cooperative Processing

Presented by: Vladimir Savic, Author(s): Vladimir Savic, Universidad Politecnica de Madrid, Spain; Henk Wymeersch, Chalmers University of Technology, Sweden; Federico Penna, Politecnico di Torino, Italy; Santiago Zazo, Universidad Politecnica de Madrid, Spain