0:00:15but that and they were mind them is the in you and for peking university was T
0:00:19uh is my for your to show some idea how to do the mobile we show such so from the
0:00:23title we focus the the first is the mobile search is a a a to different or for which research
0:00:29not the wine is though we focus on how to reduce the bad was cost a means the loop P
0:00:33three to look
0:00:34a couple encoding
0:00:36okay okay of nine
0:00:37i will give no summarise to have lot of the we search and the give some the channel to the
0:00:42regarding the mobile we should so should also like you that i'll go of this work and oh a what
0:00:47i expect that we can do to make some breaks rule
0:00:50and uh we also do you have a a a a type of the that my which was search you
0:00:54oh i at based on our experiment the result i will give the uh search you work and the we'll
0:01:00more detail but be given by the
0:01:02uh able to look at this make even more a coding
0:01:07or let's take a look at a speech or uh we try to summarise does this to have a heart
0:01:11all be sure search by Z is uh a few core
0:01:14so we have the two part of they have to put a right part for the left part in the
0:01:18region and a texture
0:01:19such a as the post
0:01:20a a local um as something like the landmark uh as that we are looking on
0:01:25uh is is somehow easy because the object use a rigid and texture
0:01:29but to for the right parts is a more difficult to a uh uh uh a a days a lot
0:01:34of the research it computer vision research has to will makes a great have to on the right part
0:01:39uh uh you you maybe you have no the google that you have that we go go dolls for the
0:01:43mobile but we should search
0:01:44i actually a product to hand that have been shifted to the uh have to part of two because the
0:01:49for the have part is a a easy uh and these possible to do that with
0:01:52you do what profit uh from using the industrial applications
0:01:57now is uh we try to come up so uh so row chan in got into the model which are
0:02:01sort the first these a really do you where is the speech
0:02:04i i think that maybe be uh is is ease on stand the mobile which are so it means that
0:02:08you you use does not camera
0:02:10you step the picture and transmitted to the server but to do some research and retaining the additional information but
0:02:15i'm used uh all three D pen with you is the use all same very very
0:02:20uh a be to a some time these uh the the bandwidth is not very stable so if for you
0:02:25to as a whole in high full told that are we are called around the
0:02:29uh fifty kilo byte
0:02:31so uh if you transmit meet a large a amount of data are then the speed that will the B
0:02:36delay they uh in you each we are have to like to he influence on the following the renting and
0:02:43the reading in the to be a data set because the the bad with the as a up upstream a
0:02:46downstream so if we spend a lot of bad with on that
0:02:49a student the delivery is the and uh at any a multimedia data sent we'll be D for that
0:02:55okay the the the two kind of the the uh use these new uh step are to to to uh
0:03:01to high degree of the that solution the first it do you we to a local to
0:03:04uh a the C hoc is the most of "'em" as a light develop by the
0:03:09uh us to have a really worth the and the the P C is if that you what you have
0:03:12a a few years ago by the same you uh a card may don't to was was day
0:03:16also the other one is the how to of the to do the compression and in level to do the
0:03:21uh something like uh
0:03:23uh i agree gay to the but three to what's you hope by the that you goes
0:03:28okay so uh uh a uh in a kind of research work the i'd uh a very promising work will
0:03:33use no compress in of what the feature to to me to to the server side but even use a
0:03:39it's very small and the good the but the is still cost around a
0:03:42two Q by to uh to it'll by to per uh we to reader
0:03:46per or uh query image
0:03:48so is uh a still um suitable for a are stable one is especially for the reality at at limitation
0:03:54because that use have to be able to do your that we you rate up or
0:03:58analysis and the retrieval
0:04:01okay out uh uh as the pipeline though the that of used the mobile about the right he's those but
0:04:06so we have a two and uh and T use the code a and and is the decoder so with
0:04:12the code and code of that we transform the lot
0:04:14uh original noble cadre of E two small mobile cover a C so with this framework a we can get
0:04:20the uh you fact you by we be to mobile search uh as a as uh the other uh on
0:04:25other hand is how to you compress the image a signature at the you meet levels
0:04:30okay oh we we we we try to a brief um the uh step are it's uh in the lead
0:04:35his uh re in the two year
0:04:37the first is the transform coding uh of the C all sub three to which what is do you uh
0:04:41do develop by the
0:04:43uh stand for a uh a you're at euros roles a group or other that you reckon component of or
0:04:48into uh image patch the so the line used the most of them are as a actually had the receive
0:04:52a lot of uh size citation asian and uh also to some work for two worlds and pack a C
0:04:57D V as a stand
0:04:59is the C hall
0:05:01okay this why is a tree structure the i the means how to compress the
0:05:05uh tree histogram coding to reduce the uh the long non parametric representation how the transmit it and the histogram
0:05:13with then
0:05:13a very small uh bit read
0:05:15and the the knots why he's the compression a look at location me for because the back a world is
0:05:20not a ways the look at will measure so how to combine the battle word
0:05:23a features a and the
0:05:25uh a spatial information to to get a some the to do some that your matching the verification
0:05:30uh uh which can improve the final performance at the server side
0:05:34okay so uh and let's the summarise the existing work there uh the the big probable and he's done now
0:05:39days the most the research will has working on the purely content best the compression then be all of our
0:05:44walk cover image signature based on which on the settings
0:05:48so uh uh uh all idea is that a a a a a shall we uh the average of the
0:05:52context you the to simple i'd the design a compression function because the with compact mission we can you reduce
0:05:57the bit rate of the us english or on the other hand that we can prove that he's been at
0:06:02the power of the the creature
0:06:05okay okay we try to summarise all up or in four to the first use the location sensitive compression because
0:06:11the we stuck on can uh location contact a mobile users
0:06:15we you can improve the compression rate as we are as as the beach or discriminative
0:06:19uh power
0:06:20the signal the this kind of a
0:06:22because for different region the that different an M are and uh for even a landmark mark that the different
0:06:27the uh uh we show appearance
0:06:28so for that i was at K show we do we develop a kind about these feature
0:06:32means that if you a difficult to recognise we can send more uh the longer to to uh if that
0:06:37it's very simple we can set the shot the so we can show at they were the scout about uh
0:06:42we should is bitter
0:06:43uh the so the wind as a a a symmetry it's because the lower they the for if you want
0:06:47to get a very promising results a you have to uh get a wrong one hundred thousand to a video
0:06:53an all that is very difficult to maintain a on the mobile side the so uh but you know the
0:06:58mobile user is with the very useful contact or major so we can make in a symmetry double cat yeah
0:07:03uh that means we can make the a very small for cadre on that
0:07:06oh about side and the but to we keep the of a large able cadre at the source side
0:07:11okay that that's why the to waco
0:07:14uh a uh so far many research what just a P D were uh uh i uh do you that
0:07:18when we coding over uh cory to the so but but in all the mobile you a mobile user has
0:07:24that to make come because you up uh upstream and downstream
0:07:27so can we use that stream to do some the supervision to smart that each of the mobile use a
0:07:32how to do that to develop a a a a a a very compact is make you that coding and
0:07:38uh a a a uh with uh and i a stream uh supervision we can get we a and you
0:07:43small and uh more robust to upstream stream reader
0:07:47okay okay that's a result to uh what the weight uh what uh what which E
0:07:51we have a a the all around it a meeting in in uh and i'm are from the pair yeah
0:07:55a and the free cost
0:07:56so uh we they are that to at the first used the highs compression rate
0:08:01uh for each
0:08:02query image we have achieved a
0:08:04paints a B
0:08:07for a image quality
0:08:09the uh the other one is the high and
0:08:11hi to want and idea like to give used on the fee go to you not region
0:08:15what it was uh what do we the
0:08:18we also prove that deployed probably the a prototype
0:08:21uh in five every including in to play area in and you all C D's by a lot of C
0:08:26and a single point them you can see the at T big i is a large area and the fire
0:08:29and see is that's small T so we try to
0:08:32get a some that you this this the to go every area to evaluate a seized
0:08:37okay okay a yeah is uh search a pipeline
0:08:40the the for is a when a mobile user enter a given location the first day
0:08:45so the location base shen we be update
0:08:47we are informed the also but side
0:08:50so that's of a side that can do
0:08:52some the look adaptation and uh transmit to
0:08:55for uh uh D the uh adaptation the from the source side in the mobile i
0:09:00and there's the
0:09:01uh user a based on the adaptive look cadre didn't K to some that coding and a D V but
0:09:06the code
0:09:07quarry no to uh a very small
0:09:09a look read to the
0:09:11uh uh remote side
0:09:12that that's a low we are ranking at in about
0:09:15okay K i stuff
0:09:16a they are the two part of the a uh uh a a a a hot is a a uh
0:09:21of side and that that uh bottom part is the mobile i
0:09:26so uh yeah is the L P V C location do screen the video coding uh it was down um
0:09:31off line at the sub to do some this special class we combine you we sure and that you a
0:09:36graphic use stands and the after the come come uh after partition we to L D V coding for uh
0:09:42the any if the coding for each region
0:09:44okay so when i you the in to the region and we update to again uh up to the uh
0:09:49location bayesian a that is such a location to a uh so side
0:09:53and a does of a the we are sent back the L D V the coding is that's on of
0:09:56that supervision
0:09:58uh the the day
0:10:00with the adaptation the mobile users can at that here and the send a compact three to the so the
0:10:07search and retrieval we have you happening
0:10:09at the server side the and find the retaining the the information
0:10:12uh related to what you want
0:10:15okay okay
0:10:16a lower uh let's give some example of the
0:10:18a some that's that the first but a graph clustering
0:10:21but that the lying the spatial cues fusion of the uh uh photos
0:10:25uh from the P in T is is um yeah it's easy to imagine that the photo uh use point
0:10:31is with uh correspond to life a little with that you tag
0:10:35right to wine used up partition the partition means a how to but it's a lot C into the all
0:10:39serious a a a sears all the small region for for each region we extract a L D V C
0:10:46okay that a uh the post and the lenny the then you process is a with dating L T V
0:10:51C by of samples the posting uh i read them to the origin or old had already
0:10:55uh key i used uh the the the target is the uh a good codebook book is expected to minimize
0:11:01the ranking lost
0:11:02uh a a i X use the ranking position function and uh that the X is the uh uh is
0:11:08the uh uh i D F distance
0:11:11i is very popular in uh a in uh traditional you for making a table now we want to a
0:11:15been my that cost uh
0:11:17to achieve that but that to figure out of the who a good codebook book
0:11:22okay uh for now uh just now high mesh we have that to do that we have done the uh
0:11:27reading partition so for each region
0:11:29we specify the leading
0:11:31so uh because the we have to take into account at both the a small sea creature
0:11:36as we all as the uh to reach people not
0:11:39so that left uh so the
0:11:42the right hard the fourth item is of medium my that cold side
0:11:45that that the part is minimize the retrieval loss
0:11:49perform the an so if you used that small
0:11:51uh look at re
0:11:54uh but the the but of a question used a how we had does the training data come from
0:11:59so uh now we we we make assumption
0:12:01if we do not do some the uh
0:12:04a small uh we keep the original large vocabulary then up of was you be good i mean the origin
0:12:09back of world
0:12:10oh oh oh we can sample as us that of the you need from each region
0:12:15we assume
0:12:16the original back a world
0:12:18yeah the a good performance
0:12:19so after uh feature action after we get uh using that small cold able a we assume a we we
0:12:26from the should be not very our away from the original problem
0:12:30so is a sort of a pseudo parry
0:12:33so for this uh now uh for these three region the we have in some holes
0:12:37uh image queries
0:12:38so we are we want tools
0:12:41we all want to minimize the
0:12:43uh mean my the class
0:12:47a yeah uh here is the us the i mean of that sample a lost to post uh we stop
0:12:52was mess or
0:12:53this part is ever weighting and this is part the ranking function
0:12:56is that a is uh explore a plane show function
0:13:00oh is like the decoder sickly creatures
0:13:02so we want to minimize the sum of the law
0:13:06the uh as is is uh for each region we have the in sample cory so we give the summary
0:13:11to and two to at you find all overall cost
0:13:15of the net be table and do
0:13:18okay the fun we we fight we trying to find out uh find out the best cold the that in
0:13:23practice we used the a greedy search a greedy algorithm
0:13:27to to uh
0:13:29uh to select the best model was that by uh one step by step
0:13:35yeah yeah is still waiting uh are a waiting a update
0:13:40so uh uh here is the compression function is uh so row minus step
0:13:45of the proposed the uh the uh a hard to a time in a to the posting we pretty fine
0:13:50a stressful
0:13:51so what the weighting error is the is nest then
0:13:56T we was no uh to the postings of it we can get of the optimize the uh we'll cover
0:14:02okay i is the uh overall of the all algorithm
0:14:05so the input is or in a battle what the features and the output the output is in the transformation
0:14:11matrix and L D V C code the we are be uh we have we got based on the these
0:14:15the uh posted it opposed us mobile cat so the yeah is a Y or loop is the posting process
0:14:24so the lost the estimation ca war the action and error waiting and of to do some the transform training
0:14:29uh that's the uh uh uh also nine of the all agree and we see called a low wrecking sensible
0:14:34country T
0:14:35for a D V to construct chip
0:14:38yeah is a result
0:14:40now uh we use comparison with a out very popular as that a of research
0:14:49this one this one is origin original bad what a bag of what so is very large
0:14:52and it this one
0:14:54uh is that a the a a after that three three if used to quick coding and this is that
0:14:58you that you goals while for a hearing a uh we actually we used does a a a uh X
0:15:03uh a from their website to do in uh
0:15:06and uh to to do the testing a data set
0:15:11the red curve is our results
0:15:12so uh and uh look at it is uh read out the uh what call and that G sorry i
0:15:18mean mean average precision
0:15:20the horizontal until is the option be so from this read that we can find a you we star
0:15:26hence speech that we have a you have a by prime using mlp precision now uh we do experiment of
0:15:31of five uh and uh all but a meeting data sets
0:15:34uh this without come from the our around a less thousand five hundred a chlorine it's that we we draw
0:15:40a we use a uh we collect
0:15:42uh a five uh five us these three hundred car is a se at the so we C to this
0:15:47so we have in total what's other the five hundred a court
0:15:53a is the some the uh not switch of the results
0:15:56so the last a that the column is the core you meet the right column use the result of that
0:16:00different the a
0:16:01uh a is each it out the the top
0:16:04the top row or is a without out uh a using all post in the interval cadre and the center
0:16:08or is that we using the origin of a a bag of word
0:16:13oh from these figure we can find that you've in with the very small smaller
0:16:16uh what kind of we still achieve better for missing uh every out that we see compatible to that we
0:16:21use in the very a large original bag or or the features
0:16:25okay uh dot now a a it's a this is uh the uh example uh you not face into to
0:16:31indicate to a is that cold a word the where is the law of batch of a what a look
0:16:36so the left the column is the query image or we can see that coke or you meet is overlap
0:16:41with that red circles
0:16:42the rest of the actual actually is that a cold or a is that a L D V C code
0:16:46a right to like it every ten read out that so the the top retrieval results so we uh if
0:16:50we you you you look at a carefully you can find that a uh are also very uh a a
0:16:55a um any uh small called walk are still overlap on the retrieval images that them actually image
0:17:02no match a cold work are a come from the L D B C code called or
0:17:07okay okay S uh another exact in that's station the full
0:17:10so that that uh that the part of the left part it L D V T called uh is that
0:17:13a serious a local patch
0:17:15and the the of that the use of chlorine the right to why i actually used uh a a works
0:17:21to the for the L D B T sat
0:17:24so uh i'll i'll main contribution is a how to do
0:17:28uh to uh to my the training uh of that that the parts the uh L T V if called
0:17:32a set from that
0:17:34uh from each region and another way used the how will take advantage of the to a coding that to
0:17:39downloading loading L L P to be say it's and a based on the download L T V say a
0:17:43and the we a though it is a uh a very small bit a very small uh
0:17:47six each chip based on the L T V he set as is L T V C set you region
0:17:51wise is we that every region
0:17:54okay K i guess some the right uh uh demonstration most original be a as you we have the set
0:17:58top the online webs that this this station to evaluate our algorithm
0:18:02well the
0:18:03for so that is uh this is that query image and the based on core you need you we attend
0:18:08the result
0:18:08is a wise tough
0:18:10a we use the google map
0:18:12a look at this a picture uh we calculate a and beat
0:18:15so for it uh of a it is a museum chair that we can get of the option it is
0:18:19that a fifty two a row of for it as it's in each car we only
0:18:24a a to the fifty two beat
0:18:27okay and an example for singapore able is tough fifty B
0:18:34uh this one is tough and come come come data set that we only need a a eighteen beach
0:18:39per or for this image
0:18:40okay so some is that we develop a back to be to compare just get
0:18:44uh the main contribution is a how to the a location discriminative what kind of coding
0:18:49and uh is uh at these uh it's east uh and T V C code it's it can make that
0:18:54discriminative the
0:18:55uh i in feature content all uh the uh the uh most important way used that we make the signature
0:19:01very compact extremely compact only ten speech
0:19:04image a
0:19:06a also he's virus kind of a because a we can stand of the most important called work
0:19:09and uh but that is a power to destroy a a way the uh that's important a
0:19:15okay okay the uh is says in in in session a set to a coding
0:19:19oh i do not know the fine and than ask question is so what's the mobile search is so a
0:19:23our argument is the mobile so is quite different to make uh we show search so the scenes a how
0:19:27to combine the be sure such and the side information then we can give V were very per missing and
0:19:31a successful for application the mobile we shows
0:19:34such so no of well now now we just use a location and information on the side of measure actually
0:19:40you use are i've id D and the some that a graph attacking there are also a a a a
0:19:44another that type of the values for side information
0:19:47so that's uh all
0:19:48a point thing to so much
0:20:08yeah on the uh right it the let there right at the centre left
0:20:12so that the home is it at i to do with the uh where
0:20:18oh you means uh you mean that is part of written out
0:20:21it that when you
0:20:22those are the return the match
0:20:24yeah yeah that top reading result he's the best the result of but uh
0:20:28uh the last to server why maybe is not a very of exact Y
0:20:33yeah i do we can would take a a couple of M I T and that T means that
0:20:37hi to retain in most are relate in the war uh a related to image
0:20:42in that top position but the we do not the cap but we have a
0:20:45whether the each position retaining the correct the result so image
0:20:49is not no we didn't use it
0:20:51because now they that we just to you the to verify with that can is K up achieve maybe that
0:20:57in the future work are you can applied your of medication it with that result saying is the regardless of
0:21:02the you image
0:21:03well as you did you could you image you consistency is well
0:21:06is it possible to get some or seekers speakers six trip for your were
0:21:10no we didn't do that to so far
0:21:12i think
0:21:12thank you
0:21:28i know
0:21:29i was interested in how many bits you need for for
0:21:32training or would only menu into
0:21:37a you see that
0:21:39of in but maybe of maybe it's don't you know
0:21:42yeah that would question
0:21:43uh now uh a now i'll walk is for of focusing on the landmark search actually the for H or
0:21:48ct we to some job right partition and a for each partition we sample some that uh some that mean
0:21:54than a mark of you
0:21:55then we sample around the sub hundred uh uh me to to do that the training phase that is that
0:22:01a case the result compact called that is better especially very suitable for uh
0:22:05search of them and an M are use the as you of the i don't know very sparse you in
0:22:10the area then the problems a one not to be good
0:22:14a bit rate you need to know how you mean the upstream or downstream
0:22:19but the don't okay that downstream stream i is a some sort of index
0:22:23around the uh wine kilo be
0:22:25one kid beat
0:22:27and the up stream is uh a power that pens but ten speech a typical T because a it's at
0:22:33per that mark image but if you want to do the uh we sure such for product or or some
0:22:39there are less texture object to then the P read we are go higher
0:22:46thank you