Damn this course looks awesome: ocw.mit.edu/courses/electrical-engineering-and-computer-science/…

@IvoFlipse Wow, it really does

Btw @Phonon would you happen to know how to compute an edge density map

> We define an edge density map for the same blob so as to complement the colour histogram. First, an edge detector (like horizontal and vertical Sobel operator) is applied to the intensity image. Then, after noise filtering, the resulting horizontal and vertical edges of a pixel are respectively quantified into 16 bins each. This will create a one-dimensional edge histogram of N=32 bins.

So what do I exactly quantify here? What decides in which bin something goes

I reckon it must be something like this:

I get that this just shows the edges (d'uh), but how to turn it into a frigging histogram :\

It's probably a histogram of edge intensity, no?

What is meant by blob in this context?

@Phonon A connected component in an image/video frame

The edge intensity, that just tells me the height of the gradients right? Nothing about their angles or whatnot

google.com/patents/US8073197

@IvoFlipse Well, it says vertical and horizontal

Which is what Sobel gives you

This is the patent I'm reading right now, which frankly is quite descriptive

@Phonon I guess you have a point there

From Googling edge histogram descriptor and opencv, they suggest to use sobel, then get the angle from the edges and take a histogram of that

Which sounds an awful lot like ICA/sparse auto-encoders

dcc.fc.up.pt/~mcoimbra/lectures/VC_1112/VC_1112_P8_LEH.pdf

Here's some Matlab code: clickdamage.com/sourcecode/code/edgeOrientationHistogram.m

@IvoFlipse Interesting. What I'm trying my head around is how to cram two completely different 16-bin histograms into one 32-bib histogram.

Makes no sense, unless a concatenation/interleaving method is specified

@Phonon I think just concatenating them

But then why make one out of it

@Phonon Either way, the patent doesn't describe it AT ALL, so I'm still at a loss

They refer to an MPEG-7 patent, which seems to make sense, but I don't get where they get their 16 from :\

@IvoFlipse Yeah, that's what patents do. They describe the work done, but they're always just subtle enough to make the product easily repeatable

@Phonon Well, I must admit they go into far more detail about their tracking procedure than any of the articles I've read so far

They just skimp over some steps, which feels odd. If you go through the trouble to point out obvious stuff, why not explain the harder stuff too

@IvoFlipse Because it's a patent = )

={

lol

I think what I'll end up with is changing my tracking code to naively only track directly connected blobs

Then come up with some metric that determines how likely it is that blobs should be clustered

Using something like temporal overlap and the spread between objects and within objects

None of the tracking solutions I could actually understand helped come up with a clear cut solution

All the ones with great looking results are littered with high level statistics :P

Yeah, makes sense. Start simple.

Well the example I showed on monday used thresholding to clearly separate between paws, because edge pixels would simply be below the threshold and get ignored. But then you have to figure out how to assign the pixels you thresholded away. This approach would be the inverse in some sense

Anyway, time for food, so I'm going to stop reading for now

And leave you with another plug for awesomely cheap monitors

oh wow, that's a pretty sweet deal

Mine is underway :)

Obviously there's some risk involved, but what's life without some risks eh?

codinghorror.com/blog/2012/07/the-ips-lcd-revolution.html

Anyway, enjoy drooling (I sure as hell did)