#apertus IRC Channel Logs

morning folks!

Good Morning!

se6astian: how can task T987 : Try running Tensorflow DeepLab-v3+ inside AXIOM Beta help?

to recognize object and beautify them, have you never used instagram filters? ;)

se6astian: ok. Let me give it try. Which processor does AXIOM has and how much main memory is available?

2 arm cores @ 800mhz

microzed has 1gb ram in total, shared with fpga

These deep learning models are pretty big not sure if the DeepLab-v3+ model fit in the available memory

but this is task really not of importance at the moment, as firmware is priority 1

I know a bit of deep learning, that is why I was asking about this task

siddhantsahu: what amount of memory are we talking here?

i think without any acceleration from the fpga side semantic segmentation would be nearly impossible

The MobileNetV2 (focused towards mobile devices)model can be used as a feature extractor with DeepLab-v3 which has less precision and number of parameters suited for mobile devices.

Refer Table 3 for the MobileNetV2 paper for model size and memory requirements. http://openaccess.thecvf.com/content_cvpr_2018/papers/Sandler_MobileNetV2_Inverted_Residuals_CVPR_2018_paper.pdf

se6astian, the menu of axiom remote is now moving when using the virtual knob, but i need some info on how it works for the real one

but if i start the visualizer and turn from default state 0 to 1, the menu entries are iterated fast till the end

let me dig out the corresponding code lines

int8_t diff = data[0] - qe[0];

yeah thats it

it works, will commit now

done, bassic version for now, soem code will be moved to other areas, like init of white balance options, just added it to visualizer for quicker results

off to bed