Current Server Time: 10:29 (Central Europe)

#apertus IRC Channel Logs

2017/03/24

Timezone: UTC


01:36
intracube
changed nick to: intracube_afk
01:47
Elbehery
left the channel
03:08
Spirit532
left the channel
03:28
Bertl_oO
off to bed now .. have a good one everyone!
03:28
Bertl_oO
changed nick to: Bertl_zZ
03:45
jucar
joined the channel
04:15
Roopal08
joined the channel
04:26
Roopal08
left the channel
04:32
jucar
left the channel
04:40
jucar
joined the channel
05:17
jucar
left the channel
06:05
dimaursu16
left the channel
06:13
d0
left the channel
06:59
Roopal08
joined the channel
07:43
Roopal08
left the channel
07:54
se6astian|away
changed nick to: se6astian
08:11
se6astian
We are starting to plan Team Talk 13 shootings
08:11
se6astian
so if you have topics that we should cover
08:11
se6astian
or ideas
08:11
se6astian
let me know
08:31
ItsMeLenny
joined the channel
08:37
Bertl_zZ
changed nick to: Bertl
08:37
Bertl
morning folks!
09:15
danieel
left the channel
09:28
sagnikbasu95
joined the channel
09:28
sagnikbasu95
Hi Bertl.
09:28
Bertl
hey sagnikbasu95! how's going?
09:32
sagnikbasu95
Everythings fine.For the sobel fpga project, I am thinking that there will be four module- a control module (which will be like a master and will monitor the communications from cmos sensor to FPGA) ; the kernel module( which will generate a N*N kernel);a line buffer module and a address control module( to ensure that correct pixels are stored in the correct address after convolution)
09:36
Bertl
the breakdown sounds reasonable, but drop the idea of sensor communication, make the first one an 'input' interface which can handle a stream of parallellized 'words' and similar for the 'address control' make that an output module which does the inverse of the input module
09:37
sagnikbasu95
like a serial-deserializer ?
09:37
Bertl
kind of, we have a massive amount of data coming from 'somewhere' and we will produce the same amount of data on the output of the filter
09:38
Bertl
just to give you an impression, let's do a simple calculation
09:39
Bertl
let's assume, we have the CMV12000 connected with 'only' 32 LVDS channels (it has a max of 64) and we are using 'only' 300MHz (it can do 600MHz clock)
09:39
sagnikbasu95
ok..
09:39
Bertl
that will produce 9.6 gigabit of data per second
09:40
Bertl
assuming that we have 12bit per sensel, this still gives an impressive 800Mhz for each input word
09:41
Bertl
now we know that the sensor can do four times as much, so getting those sensel/words one by one is not really an option here
09:41
sagnikbasu95
yes...and we must accordingly do the convolution to achieve the maximum expected latency
09:42
Bertl
so, there are two approaches now, depending on what we want to achieve with the filter
09:42
Bertl
one is to decimate the data in some meaningfull way, which really depends on the application
09:43
Bertl
for example, if we want to use it for autofocus, we don't need to store all that data which is generated
09:43
Bertl
and we do not even need to apply the filter on all input words either
09:44
Bertl
if we want to use it for peaking (camera focus aid), we need to process a subsection or a scaled down version of the input data and store/present the entire result on some output
09:44
sagnikbasu95
can we apply encoding to the data so that we know where the maximum gradient in intensity is present...so that we can apply the filter only at those addresses
09:45
Bertl
that can happen in the output pipeline for example
09:45
Bertl
if we want to do some kind of image analysis with the help of such a filter, we need to apply it to the input image (which in this case can be heavily decimated) and store the result in memory for further processing
09:47
Bertl
well, finding the maximum gradient is basically applying the filter, so unless you come up with a great idea, that won't reduce the workload :)
09:50
alexML
when decimating the image, it may be useful to do a box blur (that is, to average a NxN block, rather than taking one pixel from it)
09:50
Bertl
note that this also is a NxN kernel filter
09:50
alexML
this improves SNR, and it should be doable with reading each pixel from the input image once (and writing a smaller version of the data for future use, e.g. at next frame)
09:51
alexML
yes, but you only have to compute it once for a NxN block, not for every pixel
09:51
alexML
so the complexity would be O(n) where n is the number of source pixels
09:51
Bertl
correct, so that might be a valuable addition/option to the filter
09:51
alexML
definitely
09:52
Bertl
i.e. allow the 'input' module to decimate the data
09:52
sagnikbasu95
yes..it sounds good
09:52
alexML
for peaking on Canons, we use a 720x480 image, not the full resolution
09:52
Bertl
not sure how well that plays with pipelining though, but should probably be doable
09:53
Bertl
as I already mentioned, optimization and customization of the filter is key, i.e. synthesis should be able to optimize out everything which is not required in a particular setup
09:53
Bertl
(IMHO that will be the main challenge)
09:57
sagnikbasu95
I agree..for the N*N kernel ... i am thinking that it should store decreasing values of power of two as we move from the centre co-ordinates...with the exception that the centre row/coloumn is zero(as in case of a actual sobel kernel)
09:58
sagnikbasu95
also whether it will be zero row or coloumn will depend on the horizontal or vertical axis of the sobel gradient
09:59
Bertl
decomposing the matrix is definitely a valuable approach
10:17
Spirit532
joined the channel
10:19
sagnikbasu95
left the channel
10:46
Bertl
off for now ... bbl
10:46
Bertl
changed nick to: Bertl_oO
11:29
OnkelTem
joined the channel
11:46
dimaursu16
joined the channel
11:46
dimaursu16
left the channel
11:46
dimaursu16
joined the channel
12:04
derWalter
left the channel
12:04
derWalter
joined the channel
12:32
Tanos
joined the channel
12:47
davidak
joined the channel
13:07
Elbehery
joined the channel
13:12
RexO
And are we gonna forward this TT to early adopters and backers today or tomorrow as we agreed?
13:15
tyrone
joined the channel
13:23
se6astian
sure, why not
13:32
tyrone
left the channel
13:33
tyrone
joined the channel
13:47
davidak
left the channel
14:23
ItsMeLenny
left the channel
14:30
davidak
joined the channel
14:30
Elbehery_
joined the channel
14:35
Elbehery
left the channel
14:42
Elbehery
joined the channel
14:42
Tanos
Hello everyone!
14:43
Tanos
@se6astian can I ask you about T720?
14:48
BAndiT1983|away
changed nick to: BAndiT1983
14:54
Bertl_oO
Hey Tanos! Don't ask if you can ask, just ask away ... :)
14:54
Bertl_oO
prepend the nick and he will answer as soon as he reads it
15:00
Tanos
Ok thank you Bertl!
15:00
Tanos
@se6astian: My question is if there are any difficulties that I am missing,(something that I donât see) at implementing T720: The user creates the calibration files and itâs a matter of where to save them. So given the calibration files, at the time of creation, the proposed file structure will be created. If you want to explicitly calibrate the sensor for a set of settings then a new file will be created in a branch of the orig
15:00
Tanos
That calibration file is optimized for the shooting conditions. The way I see it is a C script.
15:01
Tanos
(I am working on creating the weekly schedule and I need some to see some sub-targets)
15:02
Tanos
*(I am working on creating the weekly schedule and I need some sub-targets)
15:09
BAndiT1983
changed nick to: BAndiT1983|away
15:32
se6astian
Tanos, correct - its not a very tricky task
15:32
se6astian
we just need to find a system that keeps everything well organized
15:32
se6astian
like when changing camera parameters we also have to load different calibration files
15:32
se6astian
if different calibrations are required for this parameter change
15:34
ChanServ
15:42
thirdpartythirty
joined the channel
16:12
davidak
left the channel
16:12
se6astian
changed nick to: se6astian|away
16:19
tyrone
left the channel
16:22
Tanos
left the channel
17:07
davidak
joined the channel
17:11
d0
joined the channel
17:12
Roopal08
joined the channel
17:13
Roopal08
Hey <BAndiT1983> , I had few doubts , let me know when u r free to answer them
17:17
thirdpartythirty
left the channel
17:27
anuditverma
joined the channel
17:36
davidak
left the channel
17:40
anuditverma
Hi BAandiT1983, I am just checking in to know the status of control daemon though I am following the repo, please reply soon, I will see the logs, thanks in advance.
17:48
Roopal08
left the channel
17:53
arpu
left the channel
17:56
Elbehery
left the channel
17:57
intracube_afk
changed nick to: intracube
18:07
arpu
joined the channel
18:35
se6astian|away
changed nick to: se6astian
18:47
Elbehery
joined the channel
18:47
niemand
joined the channel
18:53
rexbron
left the channel
18:54
rexbron
joined the channel
18:59
rexbron
left the channel
19:00
rexbron
joined the channel
19:17
anuditverma
left the channel
19:33
BAndiT1983|away
changed nick to: BAndiT1983
19:35
BAndiT1983
Roopal08, which doubts? you don't give any information on that and i cannot guess what's the problem as i've never finished my fortuneteller study
19:37
Roopal08
joined the channel
19:37
BAndiT1983
anuditverma, my regular work is requiring more time at the moment as the project there is very important, still trying to get basic communication running (it's partially done, but needs cleanup and testing), but i need some time to structure my thoughts to do it
19:39
niemand
left the channel
19:41
Roopal08
Hey BAndiT1983, you asked me to take a dng file and get it on display . You said , its already implemented on OCBackup and OCCore
19:42
BAndiT1983
yes, it's implemented
19:42
Roopal08
But I couldn't find it on OCBackup , it still shows unable to load image .
19:43
Roopal08
basically unable to load media .
19:43
BAndiT1983
you should do it in processingtest for now, i deactivated it temporarily for QtAV/FFMPEG test as i have to implement a little bit smarter recognition of the file to use the right loader, not a big thing, but my time is a little bit sparse at the moment
19:44
BAndiT1983
ThumbnailProvider is the calss which does it for OCBackup currently
19:45
Roopal08
okay .
19:50
Roopal08
left the channel
19:52
niemand
joined the channel
19:52
niemand
left the channel
19:52
niemand
joined the channel
20:13
niemand
left the channel
20:28
niemand
joined the channel
21:44
tittynale
joined the channel
21:46
tittynale
left the channel
21:46
tittynale
joined the channel
21:47
Bertl_oO
changed nick to: Bertl
21:47
tittynale
left the channel
21:52
tittynale
joined the channel
21:52
niemand
left the channel
21:52
tittynale
left the channel
21:55
onenot8
joined the channel
21:56
onenot8
hi everyone
21:56
Bertl
hey onenot8!
21:57
se6astian
off to bed
21:57
se6astian
good night
21:58
onenot8
i am already in contact with one of the member of the org. via email
21:58
Bertl
excellent!
21:58
se6astian
changed nick to: se6astian|away
21:58
onenot8
i am interested in designing the arch. of FPGA based sobel filter project
21:59
onenot8
interested in taking part in gsoc
21:59
Bertl
yep, you're not the only one ... there has been quite some interest in this task and the PWM fan controller
22:00
onenot8
i am already working on the proposal, can someone please guide me more on the insights of the project
22:00
onenot8
great
22:00
Bertl
what do you want to know?
22:00
tittynale
joined the channel
22:01
tittynale
hey!!
22:01
Bertl
hey tittynale!
22:02
onenot8
i have gone through the reference paper given for the design
22:02
onenot8
want to know more about the proposal design/outline
22:03
Bertl
more than is documented in the lab task?
22:04
onenot8
lab task ?? don't know anything about it
22:04
Bertl
https://lab.apertus.org/T733
22:05
onenot8
oh i have gone through that
22:05
onenot8
obviously
22:06
Bertl
from our side, there is no specific form for the proposal, what we want to know is how you plan to address the task, what 'milestones' you expect to happen when and what problems you expect and what ideas you have
22:06
onenot8
ok , great
22:06
Bertl
google itself has some template proposals IIRC, so checking those is probably a good idea
22:07
Bertl
we also happily check your proposal and provide comments on the task and point out flaws, problems or misconceptions as we see them
22:07
onenot8
so can i submit more than once
22:08
onenot8
??
22:08
Bertl
IIRC (but don't quote me on that, double check with GSoC) only the last submission counts
22:08
Bertl
but you do not have to submit anything for us to check it
22:09
onenot8
i don't have a new idea
22:09
Bertl
i.e. just generate a PDF or ASCII text and paste the url here and we will read it and provide feedback
22:10
onenot8
just trying to build a prototype or implement the given paper and then go for further improvements
22:10
Bertl
the paper is just for reference
22:11
onenot8
ok yes ,I'll surely get back to you after generating a minor report or similar
22:11
Bertl
the main point is to keep it as flexible as possible, while optimizing the hell out of it for special cases
22:11
Bertl
depending on where in the pipeline the filter will be used, the amount of processed data can be huge
22:13
onenot8
pipelining would only be possible in memory buffer
22:14
onenot8
convolution would be a minor combinational circuit
22:15
onenot8
I am trying to think of some optimization technique in that circuit ,
22:15
onenot8
that's exactly where I need help
22:16
Bertl
we already had some ideas here on the channel (you might read up on the IRC logs)
22:17
onenot8
thank you Bertl
22:17
Bertl
like decomposing the matrix into smaller pieces or optimizing away certain operations for a given matrix
22:18
Bertl
of course, parallelization is also a way to speed up things as is decimation of the input data
22:18
Bertl
we will probably reformulate the task description a little (or at least extend it) to contain all the input and ideas we got so far
22:18
onenot8
input data could be floating point ? right ?
22:19
Bertl
probably could be, but I guess that should be better avoided for performance reasons
22:19
Bertl
(performance and resource reasons actually)
22:22
onenot8
Yes , I am thinking of actually start coding the conventional architecture so that in between we can somehow check the performance and get the estimate of the resources used
22:23
Bertl
good idea, also some kind of testbench will probably become very useful sooner or later
22:26
onenot8
exactly and one more thing , how many sobel filters should we use , in one reference, data is convoluted with two kernels 0 and 90 degree , while in other ,I found 8 gradients calculated
22:27
Bertl
that is a good question and it probably depends very much on the application
22:28
Bertl
for edge detection (as in image recognition) separate gradients are very useful, for focus peeking OTOH, it doesn't really matter as long as you see if it is sharp or not
22:29
onenot8
ok, so maybe use build an architecture which can support both the strategies
22:30
Bertl
definitely wouldn't hurt, but maybe it is better/simpler to create variants or maybe the required resources are prohibitive and certain setups do not make any sense after all
22:30
Bertl
that is something we have to figure out during the task
22:31
onenot8
ok
22:31
Bertl
(but it is good to mention this and what options there are)
22:31
dimaursu16
left the channel
22:33
dimaursu16
joined the channel
22:33
dimaursu16
left the channel
22:33
dimaursu16
joined the channel
22:35
Bertl
I'm off for a nap now ... if you have more questions, or something to check, just leave a message here, alexML or I will answer/check when we get back ...
22:35
Bertl
have a good one everyone!
22:35
Bertl
changed nick to: Bertl_zZ
22:35
onenot8
bye bye
23:06
tittynale
left the channel
23:08
jucar
joined the channel
23:16
jucar1
joined the channel
23:18
jucar
left the channel
23:23
onenot8
left the channel
23:51
Elbehery_
left the channel