0:00:13i can much
0:00:15good morning ladies and that's judgement
0:00:17my i'm james space and i'm on his that my work
0:00:19than with P we to in mike brooks on
0:00:22that their extraction of that
0:00:23image based rendering
0:00:26my to cover
0:00:27a brief introduction of the area
0:00:29are are with them
0:00:31and evaluation of the output
0:00:32and a conclusion
0:00:34what is used
0:00:36we take just camera of these
0:00:38and synthesized new virtual views
0:00:40of the thing
0:00:42a few real world examples all
0:00:45google C
0:00:46three D a model
0:00:49and the whole time
0:00:52P C research is
0:00:53in hospice a T V and to for
0:00:57one approach to be says
0:00:59is a model based rendering
0:01:00what you take it is detail to geometric model
0:01:03i at extra map
0:01:05and generate V views
0:01:07you image is a need it
0:01:09but a high degree a much information is point
0:01:12creation model is often so and competition
0:01:16don't so approach
0:01:17would be image based you're
0:01:19what you synthesized new views directly from an images
0:01:23now geometric information is required
0:01:25and you can get a will stick account
0:01:28you do need a very large number in
0:01:32our approach is
0:01:33between these two extremes
0:01:36we have to use a reasonable number of images
0:01:40using a simple do much model
0:01:42into into a lot of computation
0:01:46and and for a jury to see
0:01:49yeah in itself
0:01:51using a bit we test set
0:01:53you have a series of cameras
0:01:54right was cameras
0:01:56a rate a row
0:01:56before C
0:01:58which is a key image
0:02:00and segment S
0:02:02sing segment rather the pixel based method
0:02:04to make it more robust to noise
0:02:06how we are assuming that
0:02:07the segments or and saying let
0:02:10is segment image we match signals across
0:02:13all of the available images
0:02:15create a disparity gradient
0:02:17which we can use to generate
0:02:18the depth map
0:02:20we can then to stuff my
0:02:23and is this to synthesise
0:02:25you virtual views
0:02:26at any point
0:02:28along the whole right image
0:02:36so if example we take
0:02:39um five input images
0:02:42you can see that
0:02:43they quite wide expense
0:02:47and as big jump change image
0:02:50if we use a method to synthesise for intermediate you
0:02:56get "'em"
0:02:56a smooth
0:02:58transition between the frames
0:02:59would no major hot five
0:03:04the scene
0:03:09the scene itself
0:03:11the in itself
0:03:12is non uniformly space
0:03:14you have objects in clusters
0:03:16throughout out that
0:03:17that that of the
0:03:18the scene
0:03:19we can represent this
0:03:20as a histogram
0:03:22all of these disparities spartans option
0:03:24we can place
0:03:26all model
0:03:27is a a is a laplace rather of and continue system
0:03:31there many reasons for this
0:03:32they in time space and the complexity of the
0:03:38if we place the late
0:03:39to minimize the error a
0:03:41on the disparity and his crap
0:03:44we can
0:03:44optimize a position of these less
0:03:47we don't waste space
0:03:50we don't wasting anything on these regions
0:03:54whether no regions of inter
0:03:58yeah the benefit of this
0:04:01is that we can
0:04:02make sure that S correspond exactly to
0:04:05the peaks which correspond to the objects and cells
0:04:08so we don't have any
0:04:10error on the whole to
0:04:12this has a new was benefits compared to
0:04:14another common approach which is that uniformly spaced less
0:04:19one not does not take into account
0:04:21the scene cell
0:04:25a second major non spent
0:04:27however find the debt
0:04:28my itself
0:04:31from all initial that map
0:04:33we have
0:04:34and signed
0:04:36and also a confidence
0:04:38in that side
0:04:40we start from close less
0:04:42because that we note occlusion
0:04:44we take a segment that
0:04:46we are compton not assessment
0:04:48we re it
0:04:50we cited
0:04:52to all new depth map
0:04:54move want to an next
0:04:55segment in this bus flat
0:04:57places laugh
0:04:58i we not competent this
0:04:59so we sets aside
0:05:01for like to pair
0:05:03as you can see
0:05:04as a meets the next level
0:05:06behind us layer
0:05:07we are only using
0:05:09segments that would call to ten
0:05:11for the next age of occlusion
0:05:13so if the next lab
0:05:14can be included so we use
0:05:17or now actually measurements for that
0:05:19classes let
0:05:20took we it is there with occlusion
0:05:23giving a some more actually result
0:05:26a finally as we finished
0:05:28through a was those less
0:05:32sect we have out
0:05:34are we can we calculate them
0:05:36using a complete data
0:05:38we can do this because
0:05:40of the occlusion ordering inherent in a like a system
0:05:43so we only need to fix occlusion
0:05:46and map
0:05:47the points at which we hit and you like
0:05:50which maximise the accuracy
0:05:52while growing now actual levels was calculation
0:05:57vol or i
0:05:59oh method of a as you are within is to take the initial set of images
0:06:04and remove so
0:06:06these are not use any anyway through
0:06:08a process
0:06:09we synthesise these use
0:06:11using the remaining images
0:06:14and then compared against the original based trace
0:06:22this all based on a tree
0:06:24we not take advantage of the
0:06:27"'kay" feast
0:06:28space less we use uniform that space
0:06:30we not take into account
0:06:32the ordering of the
0:06:34less segments as we do the cushioning
0:06:36and only using one key image
0:06:38as you can see
0:06:40as we increase number of layers
0:06:42and hence the complex you model
0:06:44the quality of our rendering increase
0:06:46how it there's a large to be a very good as C
0:06:49due to the fact that
0:06:50the lad an is in you to how many as zero
0:06:56it's such at a point
0:06:58as as predicted by an atoms minimum stopping criterion
0:07:02the not big guns
0:07:04however a proposed a present
0:07:08are optimized at position
0:07:10and not trust based occlusions
0:07:11and all that sort border refinement
0:07:13you can improve result
0:07:16firstly as much
0:07:17as a less less to to be a very good as in results
0:07:20because but each layer
0:07:22and number we optimize let positions
0:07:25secondly you can see at it class so is a much points
0:07:29this is because by a the less in the non-uniform way
0:07:32the minimum something criterion
0:07:34seems to be
0:07:35it was be
0:07:36but i
0:07:38a and this i that the same point
0:07:41it is using one key image
0:07:43if we use
0:07:44and additional key image at the the rent
0:07:46the sequel
0:07:47and most two results
0:07:48we get a two D increase
0:07:53to assess all results
0:07:55we used
0:07:56the ground truth provided
0:07:59and use the ground truth continues map
0:08:02we see this result
0:08:03as you can see
0:08:06all best approach
0:08:08which this limit
0:08:10more more some case
0:08:12would be
0:08:13take into images out
0:08:15rendering all these images
0:08:16and i've results as before
0:08:19how in this case
0:08:20there are fewer images for the initial
0:08:22it's assignment
0:08:23and if your images
0:08:25used to synthesise the out
0:08:27you was baseline
0:08:28there's a drop in quality
0:08:30but it follows a similar path
0:08:31and again
0:08:32it that a at the point predicted one min and something
0:08:36a a is them
0:08:39as there's an added point
0:08:41there's also an increase in quality
0:08:43do you the
0:08:44great i
0:08:45to receive our assignment
0:08:47do you should the
0:08:49let on the face
0:08:51a more a very space last
0:08:53and all improve for final stuff
0:08:56and thirdly
0:08:57if we use to key images
0:08:59there's a further improve
0:09:04we compare the ground truth
0:09:05so quite that but is very close
0:09:07for a for in case
0:09:08with the ground truth is a she
0:09:11for the you okay
0:09:13but everything are um
0:09:14here's an example out
0:09:16this is one of the first frames from the input
0:09:19for a challenging case
0:09:21the mean is only one point
0:09:24as you can see from the that rigid error map
0:09:26most there is are on only edges
0:09:28rather than a middle of a
0:09:32the uh
0:09:33the P lost
0:09:34twenty eight point four
0:09:38moving beyond
0:09:40the restriction of
0:09:41a right cameras
0:09:42we can use and image plane
0:09:45and in this case
0:09:46to color images
0:09:51we can move
0:09:56a all
0:09:57and then down image
0:09:58a the right
0:09:59within the image
0:10:01and then
0:10:02moving in into the image
0:10:07and the right
0:10:12are of was designed to be able
0:10:14to more what dimensions
0:10:15and further research
0:10:18in conclusion
0:10:19are are them can synthesise new views
0:10:21with low computation
0:10:23but high quality
0:10:25i like this approach gives the simple but effective
0:10:27occlusion ordering screen
0:10:29and a good approximation sing
0:10:32as can be seen by
0:10:33all place as the ground for
0:10:37the P a results
0:10:39show this
0:10:41the non-uniform spacing spacing means fewer a a a needed
0:10:44to achieve the same result
0:10:46and the minute something criteria
0:10:47can be relaxed
0:10:50and caff the selection of second ordering that that map
0:10:53refinement step
0:10:54means that we can maximise the efficient be actually
0:10:56define define a output
0:10:57with no further calculation
0:11:01i much
0:11:02or don't question
0:11:24um i was
0:11:26was wondering um are you
0:11:28uh measure components
0:11:30so or or a it's is based on a number of measures
0:11:35looking at
0:11:36or initial assignment
0:11:38and and and thing which segments tend to be missus sound
0:11:41so one of the measure we as the size of the sec
0:11:45so you more seconds of for more likely to be miss signs
0:11:48also you text within a segment
0:11:52and it's that so foreground objects
0:11:55a a of team miss assigned
0:11:57to back an objects and that the level of a possible vision
0:12:00that all these to give us a
0:12:02a a a a a small measure how come you know
0:12:15but me is cool
0:12:16one question uh maybe B is the exclamation but
0:12:19and you sure the you should do quit the real graph uh
0:12:23mention about the ground truth
0:12:25what was well
0:12:26sorry and so the
0:12:27the ground so the middle re test set a provide with the series the ground truth image
0:12:34we so you know i an we generate the already and
0:12:38that that
0:12:39and then use this to since that yeah
0:12:42so for the evaluation down the girl
0:12:44we use the ground truth provide of the test set
0:12:47to generate a while
0:12:48and then so for exact the same method is all
0:12:51the rest the vol are them but we use that that a and set of generated one
0:12:56so it's a is a measure of how
0:12:58actually the depth map is
0:13:01rather than
0:13:02and other
0:13:03red frequency
0:13:08a if there is no question
0:13:10but i think of speaker cooking