> The real reason why i posted it, is that suddenly your RSS feeds on your website work again i suppose :P
I fixed a minor bug in my Hugo setup and posted a new article, that's probably the reason why that changed. Now I still need to go fix the images on older posts. No rest for the wicked!
My current project also has ML in it; music, MIDI, programming, what's not to like :)?
I'm very much unsure about whether or not I will be able to get it to work though, I may have bitten off more than I can chew with this one.
Trying to create a very patient (and cheap) piano teacher. I realize it will never be a substitute for a real teacher but if it gets me 50% there that's already quite neat. But this is much harder than I thought it would be.
Polyphonic sound decomposition is pretty much an open problem. I have an 'interesting' idea on how to solve it. I'm sure I'll be accused of cheating if I get it to work ;)
Oh, that's true (Melodyne made a plug in that will do it, AFAIK). Can't you do something else, like optically or mechanically read the keys? Do you even need to read them? I guess you want to tell people if they made a mistake, hmm...
Something you could do would be IR LEDs and an IR camera, where the LEDs are only reflected on the shiny keys when they're pressed due to the angle.
Yes, there are many use cases that I have in mind where this is a must. In fact, if I can't crack it then it's a useless project (other than that I will learn from it).
There are all kinds of devices that could give you midi inputs (such as the 'Piano Bar', sadly discontinued), but that's the easy way out. After all, you could also simply use a MIDI enabled keyboard. But I want it to work with a real piano, as well as possibly other instruments.
The best software for this at the moment is Anthemscore.
Hmm, very true. Do you want it to not require modifications to the piano? I guess that would mean that you pretty much need to do sound decomposition...
Yes, zero hardware changes or mods to the piano. I think that would be a dealbreaker for many people. It also means that the barrier to entry would be much lower.
Hmm, yes, if you can get it to work with just sound, that would be trivial to set up. If not, I think the light idea would still be valuable, as technically you wouldn't be changing the piano (just adding a camera and light above it).
I've actually toyed around with that idea, one interesting application would be to convert Rousseau style youtube videos back into midi.
I'm quite sure it is possible but it would still require a lot of set-up. The microphone is such a temptingly simple target that it makes sense to pursue that one first.
That makes sense (although the microphone is only a simple target if you ignore the unknown unknown of deconstructing the sound :P). I'm very much looking forward to your progress!
I should set it up again and give a live demo one of these days. It's lots of fun to watch it work. At least once every week someone mails me about it. The bulk of these are wantrepreneurs who have suddenly seen the light and want to take the project over, and who think I'm gullible enough to believe their cover stories (5 minutes of Googling usually turns up a commercial operation). Can't blame them for trying to corner the market I guess.
Meanwhile, the consultancy I've been running for the last 12 years is growing up, rapidly. We just named our first CEO (other than me), who will hopefully take it to the next level so who knows, I just might have some free time again in the near future. Here's to hoping.
That would have been one case where 'eminent domain' would have been a good solution.
It's a real pity to see such an iconic building go to waste. The demolition story tries to spin it hard as though the family had no other option, but of course they did: if they were unable to be good stewards of the property then they could have donated it to the Belgian state.
But after looting it for anything valuable it is of course much better to have a white-wash story out there about how those urban explorers violated the privacy of the owners (never mind the fact that it had been abandoned for years and that there was plenty of damage from neglect alone).
Thank you. Doing anything for a long time is hard, I can't wait to be free again to do other stuff. In 2016 when I did this project I still had lots of free time, we did 11 tech due dilligences that year. In 2017 there were 18, in 2018 43 and this year we will likely end up somewhere between 60 and 80 so it long ago surpassed what one person can do.
I agree with Nico its got a lot to like, it is very approachable and doesn't imply access to a lot of hard to get/use equipment. So good stuff.
Now you need sub-sorters :-) I always kept my technic stuff sorted by various parts (large gears, medium gears, small gears, motors, axles, wheels, etc)
Another interesting project would be a 'pick a part' machine that used an arm and image recognition to select parts from bins and assemble a 'kit' (all the parts needed to make a particular thing)
The main limitations is to be able to sort into only 12 bins in one pass. If you want more detail or more groups then you need more passes. Ideally it would sort into a few hundred bins at once but I have not been able to come up with a geometry that works well. The most radical idea would be a tower of 2 meters high or so with bins spaced around it like flower petals, the air jet would have to be timed super precisely to aim the part into the right bin, especially near the bottom where the velocity would be quite high.
Pick a part is easier, the software is trivial but the hardware is anything but, an 'automated warehouse' for Lego parts would be a pretty tricky thing from a mechanical point of view.
And once you're on that train of thought, the obvious end-game is a machine that re-assembles the sets, then we can declare Lego a 'solved problem'.
Hm, never even thought that that might be a factor. Yes, it uses relays. I'm not sure how much belt travel will be off because of that, but given that it would be a constant it would likely already be accounted for a an offset to 'center of part' which is where I try to hit the parts to make sure they don't spin like mad.
That sounds reasonable, where you may have more luck with something more accurate is if you had to run the blower for a shorter time, or maybe pulse it to lower the wind speed (I'm not sure if that's possible).
The relay pulses are about 50 msec, that's mostly limited to ensure there is enough air to rapidly hit multiple bins in a short time, as well as to be able to 'pick out' a part from two adjacent parts on the belt without hitting the other two. That took a while to get right!
Have you ever posted the code? I would totally try implementing this, as the hardware work is something I can do, but my programming skills are rusty enough that I don't think I can write the code. I'd love to see other people try to build and improve on it.
No, I was planning to and then every 2 bit Lego trader and bricklink contacted me to try to go commercial with it. That's not how I envision this and so I decided to keep it inside for now. A ton of work went into labeling the dataset (many 10's of thousands of images), and there is no way I will let the likes of Bricklink and overpriced parts sellers use it to corner the market.
This is very much a mom-and-pop market, and there are countless people out there that augment their pensions by sorting Lego by hand and re-selling it.
Releasing this would put them out of business overnight, or, they'd have a grace period as long as someone does not put together industrial mechanics for it and then it is 'game over'.
