#apertus IRC Channel Logs

OpenCV has a module "dnn" which implements the most common building blocks layer (Convolution, ReLu, RNN etc.) which can be used for loading, running pre-trained neural networks (inference).

and?

More on it here https://github.com/opencv/opencv/wiki/Intel%27s-Deep-Learning-Inference-Engine-backend

For T1142 I wish to use some pre-trained neural network models for the task.

It would great over OpenCine users if they can use these algorithms.

I need user suggestion, if it is only possible to use the "dnn" module only of OpenCV in OpenCine?

of course, it uses tensorflow to locate humans in the picture and to tell us, if it's donald trump or not

you tell me, i don't have a clue about neural networks

That means once the model is build and trained on a large dataset using some deep learning frameworks like Tensorflow, Caffe, etc.

One the models architecture and the trained model's parameter.

This module is highly optimized to run on Intel CPU's

Among the two files which we get after training using deep learning frameworks (Tensorflow, Caffe etc) 1.) First the models architecture file, used to build the model 2.) Second are the actual model's parameters.

off for work, will be back later

Yes working on the C++ challenge, will send you for review once it's done.

Hi,

For streaming purpose

apurvanandan[m]: there are two option here, we can use the FT601 and have raw data and we can use the FT602 which basically creates an UVC device

in the second case we are the sender and do not need to bother with the software on the receiver side, but we are limited to the UVC formats and need to set them up properly

if the uvc can bear 422/10b, then sort of log-raw can still be passed through it, but the receiver shall be aware of it being RGGB not YUV

but the FX3 might be more flexible because of this

haven't checked the UVC specification yet, might as well be that they already have something similar defined

uvc1.5 has no mention of raw, bayer, cfa

It also has a `uncompressed` payload type

well it could have a fake debayer to yuv as initial mode, that code you already have for the hdmi

but the main plan for the USB interface is to get raw data out

mrohit[m]: you can use https://github.com/axiom-micro/axiom-control-daemon/blob/master/.clang-format for now

Hey BAndiT1983 , free to check the C/C++ challenge right now ?

like variables starting with upper-case

what should this code do? ->

file << std::endl;

AVI generation is not really usable in this form, as size calculation is very rigid

ppm is for final debayered image

and why is it in a separate line?

and messy too becouse of endl , so it shifted it in next line

ehm, don't think so, cause this looks like a waste of lines for unnecessary stuff

check your code thoroughly, as it's up to you to deliver high quality

Hi saurabh_raj :)

this hides the target of the code from observer

okay thanks BAndiT1983, if u got time also please check for proposal which i have send to team's email Id

it's not about switching to different algorithm, it's about having readable and maintainable code

i see a lot of nested if/else in the loops and so on, not really performant

okey.. so apart from the loops, any serious issues?

it's not my task to list every bit of problems there, but i see a lot of C-style code, which is not the goal for C++ challenge

why do you use so many data classes like pixelmap, raw12, bitmap?

pixelmap deals with writing ppm files, bitmap for bmp and raw12 loads the find and initializes the vectors

the structure is not clear, as you mix namespaces, classes etc., should be more clear how the data flows in the pipeline

i don't see real OOP there

struct Streams is unnecesary

usinged char should be replaced by uint8_t or uint16_t, depending on what you use

it's not my task to complete the challenge

I was trying to do the writing avi tasks but I keep on getting the error could not demultiplex the file when I play the avi file

https://github.com/BAndiT1983/OC_FrameServer/tree/dev

the main expectation is, that people want to learn new things and for GSoC it's a must, which is clearly defined by google in their FAQs

as the mentors do this stuff as volunteers, without payment

apurvanandan[m]: bonus points for LaTeX, especially on the CV :)

a link is probably the best way unless you want to incorporate it into the proposal

hey Ashu! How's going?

:) yesterday I explaine about how I wish to use GraphicMagick , is it right approach ?

well, yes and no, you are definitely heading into the right direction there

the problem I see is this:

for example for a histogram, you probably want to look at 20-50% of the image to get it resonably right

now you need to visualize it with the overlay, which in turn means additional memory access

the only difference is that it is written to the memory instead of read

but the sensitive point here obviously is 'reading' (the image data) and 'writing' (the overlay) in an efficient way

then 'simply' copy it over to the 'framebuffer' (overlay)

as the memory bandwidth is limited, this in turn means that the framerate for the overlay drops dramatically

that said, a 'render' approach would probably be the best (I think I already mentioned that but probably failed to explain it in more detail)

(this is something which needs to be tested)

but we know for sure that a 'single' write to each memory location is preferable to multiple writes

given that the 'decision' for the 'pixel color' is based on 'small' and 'local' data which is likely already cached and doesn't require any memory access, we are likely to get away without writing or reading anything twice

You are right Bertl ,it makes sense. Using library will definately require a lot of memory access from CPU side .

at least for something like the histogram it is certainly a good and viable approach

so it is likely to be a balance between our own special 'renderer' and rather generic drawing functions which can accomodate any framebuffer like setup

note that in this case the drawing should definitely be client side

(think histogram on a tablet connected via wifi)

I'm fine with that as well

haven't checked but it really depends on the visualization we want to do

you might for example want to investigate 'good' and 'easy' visualizations for the data we want to collect as well as ways to collect them efficiently

Now I can picturize real complications and problem here , It would definately require quality reading ,research and thinking to get optimise and well suited solution to this .

thanku, for explaining things in this depth . I Will come back for discussion after doing little more research. :)

Hey BAndiT1983: I mailed you my c++ challenge solution today? Did you find time to take a look at it?

s/answer/reply/

i can't understand why you store the images away, reload them again and especially the pile of streams there

how should someone be able to read this code? any quality assurance in a company would refuse to merge such code into the main branch

but those are exceptions to the rule of course

and many other vars of such kind, which makes it very hard to understand, besides that the code can be simplified by extracting proper methods and reuse them, instead of doing many if/else or switch case or even plain spaghetti code

that is called obfuscation :)

lol, yeah ... sounds legit :)

I understand BAndiT1983. I'll change the code to store the data in arrays instead of files and also follow the c++ guidelines more strictly but is the implementation of Debayer() okay or below par? should I change it to how others have done it?