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