i can much um good morning ladies and that's judgement my i'm james space and i'm on his that my work than with P we to in mike brooks on that their extraction of that image based rendering my to cover a brief introduction of the area are are with them and evaluation of the output and a conclusion what is used we take just camera of these and synthesized new virtual views of the thing a few real world examples all google C three D a model and the whole time and P C research is in hospice a T V and to for one approach to be says is a model based rendering what you take it is detail to geometric model i at extra map and generate V views you image is a need it but a high degree a much information is point creation model is often so and competition don't so approach would be image based you're what you synthesized new views directly from an images now geometric information is required and you can get a will stick account however you do need a very large number in image our approach is between these two extremes we have to use a reasonable number of images using a simple do much model into into a lot of computation and and for a jury to see is yeah in itself using a bit we test set you have a series of cameras right was cameras a rate a row before C which is a key image and segment S sing segment rather the pixel based method to make it more robust to noise how we are assuming that the segments or and saying let is segment image we match signals across all of the available images create a disparity gradient which we can use to generate the depth map we can then to stuff my and is this to synthesise you virtual views at any point along the whole right image so if example we take five um five input images you can see that they quite wide expense and as big jump change image however if we use a method to synthesise for intermediate you get "'em" a smooth transition between the frames would no major hot five the scene the scene itself the in itself is non uniformly space you have objects in clusters throughout out that that that of the the scene we can represent this as a histogram all of these disparities spartans option we can place oh all model is a a is a laplace rather of and continue system there many reasons for this they in time space and the complexity of the calculation if we place the late to minimize the error a on the disparity and his crap we can optimize a position of these less and we don't waste space we don't wasting anything on these regions yeah whether no regions of inter yeah the benefit of this place approach is that we can make sure that S correspond exactly to the peaks which correspond to the objects and cells and so we don't have any error on the whole to this has a new was benefits compared to another common approach which is that uniformly spaced less which one not does not take into account the scene cell a second major non spent it's however find the debt my itself from all initial that map we have and signed that and also a confidence in that side so we start from close less because that we note occlusion we take a segment that we are compton not assessment we re it we cited to all new depth map move want to an next segment in this bus flat places laugh i we not competent this so we sets aside for like to pair as you can see as a meets the next level behind us layer we are only using segments that would call to ten for the next age of occlusion so if the next lab can be included so we use or now actually measurements for that classes let took we it is there with occlusion giving a some more actually result a finally as we finished through a was those less we approach sect we have out are we can we calculate them using a complete data we can do this because of the occlusion ordering inherent in a like a system so we only need to fix occlusion and map the points at which we hit and you like which maximise the accuracy while growing now actual levels was calculation vol or i oh method of a as you are within is to take the initial set of images and remove so these are not use any anyway through a process we synthesise these use using the remaining images and then compared against the original based trace this all based on a tree we not take advantage of the "'kay" feast space less we use uniform that space we not take into account the ordering of the and less segments as we do the cushioning and only using one key image as you can see as we increase number of layers and hence the complex you model the quality of our rendering increase how it there's a large to be a very good as C due to the fact that the lad an is in you to how many as zero it's such at a point as as predicted by an atoms minimum stopping criterion and the not big guns however a proposed a present using are optimized at position and not trust based occlusions and all that sort border refinement you can improve result firstly as much as a less less to to be a very good as in results because but each layer and number we optimize let positions secondly you can see at it class so is a much points this is because by a the less in the non-uniform way the minimum something criterion seems to be it was be but i and a and this i that the same point it is using one key image if we use and additional key image at the the rent the sequel and most two results we get a two D increase course to assess all results we used the ground truth provided and use the ground truth continues map we we see this result as you can see all all best approach which this limit more more some case would be take into images out rendering all these images and i've results as before how in this case there are fewer images for the initial it's assignment and if your images used to synthesise the out you was baseline there's a drop in quality but it follows a similar path and again it that a at the point predicted one min and something a a is them as there's an added point and there's also an increase in quality do you the great i to receive our assignment and do you should the let on the face a more a very space last and all improve for final stuff and thirdly if we use to key images there's a further improve and we compare the ground truth so quite that but is very close for a for in case with the ground truth is a she for the you okay but everything are um here's an example out this is one of the first frames from the input for a challenging case the mean is only one point four as you can see from the that rigid error map most there is are on only edges rather than a middle of a and the uh the P lost twenty eight point four moving beyond the restriction of a right cameras we can use and image plane and in this case to color images we can move a all and then down image a the right within the image and then moving in into the image but and the right are of was designed to be able to more what dimensions and further research in conclusion are are them can synthesise new views with low computation but high quality i like this approach gives the simple but effective occlusion ordering screen and a good approximation sing as can be seen by all place as the ground for the P a results show this the non-uniform spacing spacing means fewer a a a needed to achieve the same result and the minute something criteria can be relaxed and caff the selection of second ordering that that map refinement step means that we can maximise the efficient be actually define define a output with no further calculation i much or don't question questions i um i was was wondering um are you uh measure components so or or a it's is based on a number of measures um by looking at or initial assignment and and and thing which segments tend to be missus sound so one of the measure we as the size of the sec and so you more seconds of for more likely to be miss signs also you text within a segment um and it's that so foreground objects a a of team miss assigned to back an objects and that the level of a possible vision that all these to give us a a a a a a small measure how come you know thing but me is cool one question uh maybe B is the exclamation but and you sure the you should do quit the real graph uh mention about the ground truth what was well sorry and so the the ground so the middle re test set a provide with the series the ground truth image um we so you know i an we generate the already and um that that and then use this to since that yeah so for the evaluation down the girl we use the ground truth provide of the test set to generate a while and then so for exact the same method is all the rest the vol are them but we use that that a and set of generated one so it's a is a measure of how actually the depth map is um rather than and other red frequency a if there is no question but i think of speaker cooking