MemoryAvailable[] returns the available memory reported by the operating system. Any way to mimic this with a user defined function in older versions of Mathematica?
I'm hoping you can help me again, this is driving me mad!
I'm creating a MatrixPlot and then exporting it to a file, like so: i = MatrixPlot[m, MaxPlotPoints -> {360, 360}, Frame -> False, PlotRangePadding -> None]; Export["mp/" <> f <> ".mp.png", i, ImageSize -> {360, 360}, Alignment -> {Left, Bottom}];
but for the life of me, I can't get it aligned in the bottom left corner of the exported image.
I guess I'm doing something stupid, but what?! Any ideas?
I couldn't seem to understand, from couple of days, my code are saying ParallelTable::nopar: No parallel kernels available; proceeding with sequential evaluation.
But never happened before couple of day, if I'm not wrong I hardly changed any settings
@C.E. Cool question. My first thought is to look at the construction notebook for the YOLO network on the neural net repo. I'm outdoors today but it's a cool idea
Do you have a dataset?
I wrote a segmentation network in a recent answer, you could almost definitely repurpose it
That particular net is very fast, so it would potentially work for your end goal of active video AR
(I'm thinking of segmenting the rectangle that those points are the corner of)
Here is a rudimentary, but very fast, neural network approach.
What we'll be doing is generating masks within which are our circles.
First, we'll come up with a slightly faster way to make the images. RandomImage will already generate noise quite simply, and then we can Blend your circles into ...
@CarlLange Your results in that question are very impressive, especially that you got such great results with such little training time.
The important thing, in order to be able to find the transform for AR, is to find $n$ points with great precision.
Segmentation may not be the way to do that (?)
I did try to train Yolo and it looked like that, so yeah, it seems to work pretty well. That's just with approx. 100 photos, ~1000 annotated Pokémon cards.
Maybe it can be that basis for something, although I wouldn't know how to modify it.
Re. training data for this problem, I haven't done it yet but I can get a lot by using the algorithm that I wrote about. I can take a thousand or so photos and automatically annotate them. Then, if that works, I can take more difficult photos where that algorithm doesn't work and annotate those manually.
But I can't use those for transfer learning, I'm pretty sure, because the problems are completely different. There are plenty of networks for detecting objects that should be suitable though, but I don't know how to adapt them. (Only thing I've tried so far: Stick a couple of linear layers to the end of VGG-16.)
@CarlLange Enjoy the time outdoors, I'll tell you when I have a dataset :)
@C.E. I would bet that you could take NetModel["Vanilla CNN for Facial Landmark Regression"], change the LinearLayer to 8 and the Reshape to 4*2, and it would start to work with some training.
only thing is it's for very small images, but it would be a proof of concept