Separable Filters and a Bauble - Computerphile 

I've seen this done with a discrete cosine transform (which I believe is a special case of a Fourier transform). It was actually pretty neat, and SUPER fast.

AFAIK convolutions become multiplications in frequency domain. so you take the FT of the image and the FT of the kernel, multiply them together and then take the inverse fourier transform of the image

@@calaphos Essentially yes. You would need some padding and cropping to not get cyclic behavior (left side blurring onto the right side and vice versa, and same for up/down).

I'm no expert for the FFT, but it's "just" an algorithm for the discrete Fourier transform, originally only for cases where you dealt with N=2^n values, that is more efficient than the classic approach. The classic DFT algorithm requires a number of steps proportional to N^2 whereas the FFT requires only a number proportional to N log N. This is the same proportional difference, for example, between the average timings of bubble sort and quick sort and does make a huge difference for large N. Also, the Fourier transform is incredibly useful, image processing has already been mentioned, but also e.g. digital audio compression, statistics, cryptography, number theory to name but a few.

Self deprecation.. a fine form of humor ;)

I just wish he'd actually do something about it.

@@Anvilshock His camerawork is fine.

@@General12th No, it's idiotic, hyperactive, nauseating garbage.

32 butterflies liked your comment! :)

It would be a short video. After a FT, a convolution is just multiplying. Taking the DFT(discrete version of FT) takes just as long as a single filter the size of the full image, but you can then do as many convolutions(of any size) as you want. FFT is closely related to jpeg compression and you see the same types of artifacting, but it is fast and allows for unlimited conversations like a DFT. TLDR, that is more a topic for numberphile. The convolution is so easy in the frequency/spatial domain that it is trivial, but the reason why it works is very math heavy.

on my machine, the Windows 10 photo viewer takes up over 600MB of RAM just while viewing ordinary jpegs...

Candy Crush Photo Viewer

it's because the new photo viewer is a "Universal Windows Platform" app which really means it's written in JavaScript and has to load a web browser to view your photo LOL

take a look at the program folder, a huge amount of files and dlls, remember that it is necessary to configure the programs folder and give access to you (user) to be able to see the files contained, a lot of unnecessary. I'm almost done with my 15mb photo viewing program that supports the formats written by me .tga, .pcx, net pbm and some other .jpg, .png, .tiff, .hdr, .psd, giff libraries. ..

thats so insightful, you must be python expert!

Just a friendly reminder that pypy exist.

Anyone who complains about python being slow either has never used python or has tried a really naive implementation of a solution to an expensive problem in pure python instead of using standard python tools like numpy or pandas which are implemented in C. You can see in this video that he's using numpy. Python is fast if you use it correctly.

@ebulating was your python implementation using numpy or pandas?

@@jackeown "Python is fast if you use tools that are not in Python"

Ah sorry, bear with me >Sean

Waiting for that code too! Come on Mike >:(

Chill it's just 30 minutes since the video :P

@@DroidFreak32 well, i assume it has been edited for some time between filming and posting. even if that's not the case, he had 5 minutes already

not Mike's fault I just forgot to paste the link in :) Sean

It depends om what you want it to do. You can ignore those pixels and get a smaller output image, you can pad the edges with zeroes which will result in a dark edge around it, or you can pad the edges by repeating information in your input. This could be repeating the edge pixel, mirroring pixels around the edge or wrapping the opposite side of the image around.

image rotations arent free.

@@styleisaweapon In this case a rotation would be pretty much free compared to doing a convolution, especially if your kernel is decently sized.

@@alcesmir Saying it doesnt make it true. A rotation forces two different access patterns, one with short stride, one with long. Doesnt matter which one is the read and which one is the write. Stop trying to get a free lunch by hiding the work in an abstraction trick. The abstraction also has costs.

cache misses have costs no matter how you try to disguise them or hide them or pretend they arent there - the proof is that you are trying to solve the cache miss problem - by doubling the number of them

@@styleisaweapon My point is not about reducing cache misses. My point is that doing the rotation O(width*height) is cheap compared to the convolution step O(width*height*kernel size), especially for large kernels. I agree doing the rotation is pretty useless in this case, but the argument is that you're just introducing (potentially more) cache misses somewhere else, not that image rotation is massively expensive.

I had to manually select 360p for the video to load.

And no speed options

watching it in 360 because I can't wait

@@sn0opyKS but the "Extra Bits" video has HD as an option already

Maybe a high-quality low-quality joke?

sounds like my camerawork

I own several, they are one of the most comfortable laptop keyboards and they look nice with their classic black monolithic design

Wait a few hours EDIT: Oh I got the joke

Is this supposed to be funny?

360p tho

10:20

depending on how you view it, it's either interpolated or downscaled #nerdoff

Panasonic LX100 >Sean

@@Computerphile Looks quite similar to the Sony Alpha 6000 series. Thanks.

Python is inded fast enough for 99.9% of use cases I've found. In this case, you need a bit of a workaround. It kind of depends on what you mean by python being fast here. The Numpy approach would be fine if you vectorised it all, but in some sense by fully vectorised what that means in practice is the heavy lifting is done almost entirely in C behind the scenes. This isn't a complaint about python, it's just how it's designed. I tried a partial vectorisation, but I left it at that because using a full vectorised approach (or indeed the convolve function in numpy) would have somewhat defeated the object of the demo. Similarly Opencv just uses an FFT, which again wouldn't show what I was trying to do!

Oh come on, Python is inherently *extremely* slow, and also constrained by its infamous GIL (global interpreter lock). There are indeed many things you can do to improve performance, from cython to pypy (I love pypy myself!) but pure Python is actually much, much slower than C, Java, Julia, Haskell etc. What helps is that part of its "batteries included" library (BLAS routines etc.) is actually written in very fast C, but if you have to write preformant code yourself you should be ready to use its FFI...

(and then it's not Python anymore)

n=1?

Maybe watch the "beast gpu cluster" video >Sean

photopea

What *is* a scripting language? (That's a real question.) If you want to define that with respect to the way it's executed (with an interpreter, with a compiler to machine code, or with a compiler to byte code), that's actually incorrect. And actually, the notion of "compiled language" or "interpreted language" doesn't really make sens. There are C interpreters and and there are python compilers. And neither the language itself nor the way it's executed decide its performance. Compilers have all the time in the world (kind of) to analyze the code and generate the best possible machine code. Interpreters (and virtual machines) can collect statistics and perform the optimizations on the fly based on the actual distribution of values that your program runs on. There's no clear winner as of now.

@@Ceelvain obviously by "scripted" he actually means "interpreted." One of those situations where the person knows just enough to think themselves well versed while actually tricking them into being publicly misinformative.

Not odd at all, very typical actually