and you hello and going to talk about the signal then an estimation of stereo party and optical flow we have a mean a curve video set up it means we have to calibrate it and synchronise came observing a scene and producing a stream of is is is there are images and the task is to compute uh this this party map can is there are sir the middle i i can show just uh the computed disparity map between uh between they're a pair and thank you and the optical flow maps between a a consecutive images and together with the calibration this kicks that's that's we D scene so so you you flow it's emotion you of it means that for each reconstruct point we have that to it that it's but C so how does it work with a a simple geometrical uh we have a you can cameras and to a point in three D space this point project to the images and having a the corresponding all these two images we can you can reconstruct C point white a relation that and this this point moves in time T plus one when not or of creation to gain project you me and the from the corresponding from these part we can compute we can reconstruct it one in three but we need to know that points belong want to get or that uh these course and therefore we need to compute also the correspondence i in D in no uh it of the images which is in fact you can flow uh so that that the the task is the plot we we we are given we are and points yeah the the point X time that's time one for and we should compute the one at time T possible one in the next frame where is really point means we we have to compute uh this optical flow and uh this topic of flow in right image in class or or we can compute yeah to of no into a in that image and the just in and in the in the second frame or or we can compute as is per uh optical of in the right frame and the disparity i in the in the second frame which means that these problems are couple that we can a that they usually to each other a you because we have more constraint that the un use task because he's user uh before or well explained and the output hmmm uh show you the results that that you can have tuition well well but you can expect uh so uh this is an input image let image each or is also right image but i don't show it it looks it looks similar this is the disparity map or or to it where my course are close to the camera and the black is and that are on mixed pixel and uh this this is this is that and this is the motion map this is a horizontal component of the optical flow and vertical component of of the call for what gain or coding uh new means that right and down motion yellow or right means left and a motion i and like the video uh you can see D uh how to this party and motion yeah walls well so this basis the in fact this this team of to presentation of the see the so everything single i see now or one on a correspondence problem the correspondence problem is one of the fundamental problems in computer vision and approach leads stop button "'cause" we have a lot of ambiguity but what what else uh i for constraint of course and the new you we use a constraint that each there and optical flows some spatial smoothness because never break is to have similar parties it's two but not at occlusions or object and function boundaries and then process because and of the disparity and optical flow changes abruptly except for or or your quick and about to pound there some solution in the age sorry is to use an explicit regularization uh this this so it's typically it to M R F and partition addition relation uh which is and to because it very computationally in in dense and moreover it produces it's very uh a it it might produce the artifacts which were cost them that were prior model so over the we data and this is not suitable for some applications i was to be we can there are other approaches which are discriminative but it's which just keeps us gives up but the i make use part of the solution and finds only the on a bit used to use far and we we we for to do to in this way lee back or for this purpose we you know that it test seed growing technique uh the basic idea of this going to technique is that we have a set initial correspondences so called and uh uh the these the other correspondences are are found in a small neighborhood around these C and then these you correspondences are you seats end the that this way the the the growing process continue and we have a recently developed a you growing stereo or not not a not but only mister do you some of these that is was published in but mean and a we view be scrolling not work my such way of between uh a of energy minimization that centre complete complete D local metal because the neighbouring structure of the solution is not leak more completely it's uh the the growing process rick is is the solution implicitly a the other uh_huh dissolves a robust against the initial seeds and it's very fast do to to search space reduction action it's and power three goes to and about to where and power to is the the and and and it's quite is that is the size of the images so that that the exhaustive this part space is of this i and the the the work of the uh and that it doesn't produce fully then so results we don't match all pixels in the scene but only set uh some subset which is for for for many applications that is factor so before i'll explain the seen how how we grow to see the why we you if how the uh stammer or row wink our work uh so is there and we have only two images left and right and but say this is the correspondence seat which are the the the C can be obtained by matching a distinct uh have for distinctive if you image features and uh the growing process finds the correspondences in the in the neighbourhood of the C so would it's a four to the right uh it performs the local optimization of the image correlation i still because the pixel which matches the best and if the the correlation is about a trace row then this is accepted you match and uh and uh uh where same and okay the same and these are you match found these matches because seat ten okay a process with pete uh in you can see it for this is as there are there are with the seat and the disparity map scroll so from a single seat you can grow to be a large so uh to up to uh with with this with the scene flow it's it's pretty much yeah similar we have to grow simultaneously disparity map and optical flow so we have well us there oh to there there is that time one and uh time T plus one and the seat it's not a pair of images at at a pair points a course for it the see is a correspondence of for or so it fully determine the that the the local scene flow and uh we have given it is point to map at time T one from from stereo matching i ching for from a previous frame and then the seats the seats are they and here we used in our implementation by matching harris or and low look look as gonna tracker of harris point to obtain a local of the "'cause" and then we we the growing process is the same so it a a it looks in the neighbourhood of the of the initial C the it is like a local bic you can five speed in the paper but the it locally fines locally maximise the correlation the correlation measures uh the similar to all all all all three correspondences and if the correlation is about a racial exceeds X threshold and and you match just found before a a a a a a it covers this what and that's it it's this is that pretty straightforward extension all of these we a of this uh there are going to work but it it it works quite nice so well uh for for the results we perform a think that the ground two experiments want to have a way to uh the performance of the outboard in so we seem to ties the a or playing with the tech texture play which works moving and we we can at the a what with with with noise and we also the D what the seem like we texture and that the the the conclusion of this expert experiment is that the similar in is estimation of joint formulation of optical flow and disparity uh a house a lot and how and uh uh the other work is better than independence dependence on top to cool well you can find it doesn't the paper is i'm not cool in uh some some under real results uh i again the same left image disparity map and to a motion the horizontal and vertical components of the of though you can no the is that the was between objects are what and and they are no that no smoothing got up there are no uh even for a for for you objects which are close to each other different that the different motion that the uh the are is are not confusing there are some and that we that the solution is not fully that that but the i i believe that for many applications is now another example uh the i was clapping can and C D uh colours so and that are and under example is uh moving camera that was predisposed was is that the camera and that the rooms moving we was mounted on a and that may we can arrange and strolled a tab street so we can a the result of the T O what you can find here so the the cameras with be quite right and that that the C C is complex they are many uh a menu objects in three D the various motions but has to hence the cars and the and in the they are shot boundaries a is just a a you objects are and nicely lead these thing which from from the background you with in in that and motion so uh the conclude a uh i would summarise the problem so the proposed that work it a a large displacement between frames more we have more than for a certain pixel in the last a stick for forty which is which is a disaster because it four uh a or was which you for of the variational optical flow and of because the S you in and is not motion uh that or bandages is that i the the which in boundaries are what reserve uh as a problem for hmmm i wouldn't with strong relation they tend to smooth out clues about every is but to maps are for late like what we are the the but are we have a better result than three by friends there one for based thing optical flow yeah you could you could see some flickering is true but the it's a we much better than computing frame by frame and in that the other i think is that uh the other lost it to our implementation is how to sell to team still are like matlab it's france in this or a solution about five or what i have a five seconds it up with that normal what to P C and it means that there was not a significant extra post with respect to a a single step um uh uh uh the reason is that a low low covered make a as a low of with a complexity the the search space which is to be but the correspondences are sort is and how to five for each pixel you have to find that this partly or something vertical to control three uses two and squared the quite the size of the image and that well because that's results are not lead that spot seven we had a nest and version of this song one disciple a point in that a a in the by i wine uh for for for processing the store with with also motion prediction and we have a where in this if you got a paper we have uh experiments a person with other state-of-the-art my thoughts in a in a scene flow special the for the topic of okay thank you can have some questions yes uh i just one thing did you have bet against uh we did but it it has it's and are if you so that you would i come undone god now in this case we didn't comparing in a better T as it because uh a the our our or move does not provide fully dense results and it's not comparable a few it you compare only some part of the solution and to the this party everywhere so it's it wouldn't be fair compare and other question the man a question you so for for is uh the the results are not then so but the if somebody needs to be to have some and we don't to have a any idea of to make it a a more dense okay the it that the uh miss some sub or it's how to control the trade-off between density and or right uh we break her in in for our applications that the that there are less errors and oh it's also has density which is natural i have but you can you can slide control if you relax the artwork and to match then uh the uh then you are getting more more and more illusions send then in a do you want to knit some okay course you can do some processing a some some are are a lot of use you if you know what you are looking at you can interpolate place some some it you can learn some so two for one uh this is another way and T another another option is of course two to incorporate the primary goal you mice similar or to the the you to my station that okay there is no question ms a course again i