It works!!!!! WOOOOO YEAHH!!!

Neat ^^

This might sound super salty and unnecessary hateful, but how much actual code work is done by you?

Yay! Congratulations.

@daintycode it's a fair question. In regards to the object detection. Most work was the dataset.

1. annotating images.
2. Artificially inflating the set by scripts which transform, crop, decrease image quality, cutt&paste annotated objects to different pictures while maintaining and/or generating the correct annotations. This took quite a while.

Not in this is the android integration. Running object detection on frames pulled from a recorded (ffmpeg on android) video and writing these into a file.

Next step is to train a classifier on the generated output directly on the device and since TF only has inference on android (there are repos which include part of the training function directly in C++, but I'm no good in that language and I have a time constraint) I've had to include deeplearning4j on android. Unfortunately their arm64-v8a snapshots don't compile. Essentially, a good deal of time went into smoothing out library comparability issues.

@ewpratten https://goo.gl/6Frcdd

Latest changes are not live. It's a pretty large repo since both TF and TF-Models are submodules of it. I probably don't need everything in there but I'm too scared to delete random parts only to find out I require them later.

There's also the datasets (thank god for git lfs) If you're interested I can hit you up once it's done!

@ewpratten scripts are in python, android part is in java.

@daintycode
Do you like crashing people happiness. Maybe it's his first time doing this kind of project. Or maybe his first time coding.
So stop doing that 😑

@Abyss it's cool. besides, it's not like 'lines-of-code' is any kind of indicator of how good or bad an engineer I am (although, admittedly, it took longer than it should have because of a fucking retarded error I made ^^)

What is coding without making retarded errors? @polaroidkidd

@Abyss dudes like you are the reason i wrote that it might sound wrong.
I really like this, saw his first tries with it and upvoted it, there is no feel-crushing involved whatsoever.
I just wanted to know how much code stuff like that needs, because i myself never got into ML but love the results and I really like to reinvent the wheel.
My question was more meant like you asking someone how much banana they put in a milkshake. Pure interest.

And I liked how detailed he responded.

We cool?

@polaroidkidd I told you that you will do it!

Did you use a library or API for the detection part?

wow, 100% sure? what happens to that 1.00 if you turn the square 45°?

and, oh yeah, congrats man 👏

@daintycode
Yeah we cool

@daintycode goats dude. Getting this part to work honestly isn't that much work. I took 5-10 videos of the shapes I various places and backgrounds (each vid was maybe 10-20 seconds long) and pulled 20 frames from each. In the end I have about 200 images on which to train it. In the thesis I'll have an "installation"part in the appendix. If you want I can share it with you once it's done.

@RodrigoF I used the tensor flow detection API.

@vhoyer I had totally not thought of that!

No big deal
https://play.google.com/store/apps/...

@HTMLspecialist if you want it to detect custom objects, here's a tutorial -> https://goo.gl/JR7Drv

Have fun!

@ladiesman nope. Tensor flow lite

@polaroidkidd Now use the model in tensorflow.js and anybody will be able to check it out on your website.