Density raises $51M Series C led by Kleiner Perkins
density.ioI have to admit, more and more I read about things on HN and just reason "ah this is tech beyond my grasp, and it makes sense to someone somewhere but not me". Lots of developer tools for languages I've not heard of or libraries that seem so remote and arcane I wonder who is using them.
Now in this case, it is solving a real problem, with a clear way to make money. Go Density Go.
Thank you. Fwiw, our eng team has deeply benefited from all those dev tools and libraries. The stack to result in real-time, accurate count is nuts. I think one of our devs, Gus, was on the thread. Maybe he can weigh in.
Right now our production backend relies on Postgres, Django, Kafka, Docker, Nomad, Celery (yes, those queues probably should be folded back into Kafka eventually), and lots of other open-source tech. I'm not sure if I should share much, and can't talk in depth anyway, about the embedded stack. And then there are the tools for developing the algorithm, and custom systems for assembly/test/pack in Syracuse.
Our web application is mostly TypeScript, React, RxJS.
The hardware is yours?
Yes, designed in-house by some very talented colleagues.
Nice job, that ain't easy.
Do you worry that you're taking money at a valuation that's predicated on artificially and transiently elevated demand? It doesn't sound like you have a business compatible with the VC hyper-growth model, outside of the present extraordinary circumstances.
It also feels intuitively like many person counting use cases are satisfactorily addressed by far simpler and lower cost heuristics such as, for example, Google lining up device 'throughput' with lat/long and business records. In fact, I just used their estimate to time my visit to the DMV and it worked perfectly.
Admittedly, it may just be that I'm not creative enough to envision sustainable, valuable use cases for this. But if you've been around 6 (?) years and just took $51M in Series C without being significantly diluted, KP is presumably assigning you at least a $250M valuation and possibly a lot higher. After accounting for terms like e.g. liquidation preference, the economic picture of Density could easily be a house of cards.
Lower cost systems like Google's Popular Times, are not as accurate as they seem. The next time you see a graph on google, you'll notice there's no y-axis. It's a clever way to look like actuals when it's relative to itself. Also we have units in many GPT locations. They're not accurate. On the vc returns thing ¯\_(ツ)_/¯ We'll see.
How does the accuracy of your system compare to ARKit’s people/skeleton tracking with the LiDAR Scanner in the new iPad Pro?
You're asking good questions about raising capital but the outcome was a rare circumstance where all cofounders, previous investors, new investors, current employees, and the exec team are thrilled. There's a lot to do but the round was right-sized as it was the same amount we planned to raise pre-pandemic.
I remember when this was a Show HN, it was my first comment to say that the homepage didn't work for me in an old version of Firefox (and the developers reached out, thanked me for the report, and fixed the problem).
https://news.ycombinator.com/item?id=9878597
Now so many businesses needs to count capacity, and they're perfectly positioned with 5 years of experience in the space. Amazing foresight (and/or a bit of luck).
dtertman! Those were awesome questions (same cofounder here). I read a thing a while back as a startup, your job is to survive long enough for the market to need you. Great talking to you 5 years later.
earlier thread - 2014
From density.io:
> Unlike other systems, Density is anonymous at source. We can't fingerprint or track you if we wanted to.
I wonder how true this statement is, given how accurately we can now identify people based on just their walking gait – see "Human Gait Recognition Based on Gait Energy" paper[1].
The 'sensor' view on density.io that attempts to illustrate this point appears to convey just as much as the sample images in the paper.
We get this question from time to time. It may be possible with some systems but it's not feasible at our resolution. You can see more of our depth data here -- https://medium.com/density-inc/ai-is-not-magic-its-manual-la...
I'll have to take your word for it – I'm curious if you have tried it though. I can certainly imagine that it might be feasible (even at lower resolutions) with enough stirring of the 'linear algebra pile'.
Not meaning to rain on your parade though – congratulations on a great idea!
No rain! It's a great callout. We'll see what we can do if for no other reason than to understand risk.
This is a great example of taking an innovation and positioning it as a solution to a real, painful, and timely business problem.
Many startups in their place try to come up with some original messaging and as a result overcomplicate everything. Density keeps it simple: "Count people."
I've lost track of the number of times we've said, "well, we count people" and thought... god this must be boring for other people.
I don't understand why this can't already be done with CCTV cameras. There's obvious privacy concerns, but we already have CCTV. We are already recorded. Why do you need a special device when you could just count people on camera.
Oh man, I remember about 10 years I got so annoyed with my buys-ass gym and having to wait in line to use a machine that I came up with the idea to track real-time gym attendance and machine usage. It was going to be called WaitLift. It never went further than that awesome name.
That is an exceptional name. If you still want to build it, we have a real-time API. Love supporting apps :) (Density cofounder here)
I went further, setup a landing page and sent out some customer validation emails...most responded that they didn't have the problem.
Ultimately I didn't pursue it mainly because I had no real passion for the idea itself; it was a problem only when I was traveling and staying at busy hotels with many other business travelers wanting to use the gym at the same time.
NoLinePump, ZeroLatencyGains, etc. A marketing person would have a field day naming your idea!
I wonder how much more accurate this solution is compared to a simple IR beam across the entry. Is the extra HW and SW complexity needed to implement a computer vision solution worth the increase accuracy? I suppose the answer would depend on the use case. Mom-and-pop shops might only need an IR beam, if that, while megacorps that optimize costs down to the penny might have use for exact customer trends so they can decide how much human workforce and/or automation they need.
Basic IR beam solutions cannot tell the difference between someone going in the store vs someone going out, so keeping a real-time count of occupancy is difficult (or impossible). They also struggle with two people going through a doorway at the same time (counted as one entry vs 2).
A two beam setup would allow differentiation between entering and exiting of single clients. Then the question comes down to how often do 2+ people cross the beams at once, and does that have enough of an effect on the aggregate statistics, if it was taken into account (e.g. count each crossing as 1.1 people, if 10% of beam crossings are 2 people)?
Can confirm
> We reasoned, if a solution didn't already exist — which surely it must — it shouldn't be more than a weekend project. How hard could it be to count people?
Not sure if this is really what happened (no reason not to trust them; but sometimes founders' stories are rewritten), but it's just so cool to start like that and become a large, well funded startup.
This one's legit (cofounder here). We honestly thought it would be a side project. We originally used a raspberry pi and two TP-Links to recreate a router (https://www.youtube.com/watch?v=yRGa9-QUDWo&t=7s). People are really hard to count.
Great team with an excellent product that was able to re-position extremely well during the current COVID era. Hats off to all at Density.
Worked with this team ~8 years ago when in college at my first startup and they were a consultancy. Smart people who I learned lots from when starting out and always enjoy hearing their continued success.
On my street in SF that has a dead end, cars keep coming up looking for parking and on weekends they usually find nothing. But they get stuck turning around and cause a mess.
Could I use Density to count cars and display number of empty spots at the entrance of my street?
Can't count cars at the moment, unfortunately. But there's some cool tech from Flocksafety and other camera solutions.
Obviously impressive and good timing. Congratulations.
But, I'm afraid this is what popped into my head. It must be me: https://www.youtube.com/watch?v=PqKmb7SSpTE
Gold. Proud to be affiliated.
A pretty cool product and very useful! I hope something like this becomes super wide spread (maybe with competitors as pricing is a bit wack) so we can always know how busy our local stores etc are before heading out.
If it becomes super wide spread, we'd make it low to no cost to provide wait times to end users.
this is an awesome product. I could see tons of businesses using this to reopen safely during the pandemic. +1 for making it anonymous anti-camera
Aside: Haven't seen "Made in USA" in a long time, refreshing to see that Density hardware will be manufactured in Syracuse, NY: https://www.density.io/hardware/
Anyone know what kind of depth measurement they are making? It looks like some kind of radio widget
Density (software) engineer here. Our sensor uses active infrared lasers to read depth, and it classifies "persons" with a deep-learning algorithm that runs on the device.
Why not just scan for MAC addresses? Legal reasons? Can't you black box something like that and not safe the actual IDs but use them for unique counting?
Most OSes and devices randomize MAC addresses during scans to prevent this: https://en.wikipedia.org/wiki/MAC_spoofing#MAC_Address_Rando...
It's also not guaranteed that every person is carrying a phone or that their phone is turned on.
Very cool!
Thank you, I agree! My focus is mainly on the parts downstream from all of this, once counts & count deltas can be uploaded to the cloud. When I joined, the focus on advanced hardware, anonymity/privacy, and edge-computing all struck me as very forward-thinking, and it's been really cool to see the product mature and find its market in the past couple of years.
there's a grocery store equivalent in the UK called Crowdless
That seems like a decent solution if you don't care about any sort of accuracy, and are ok with very rough estimates (which is certainly reasonable for some use cases).
Who would need exact accuracy? Why won't people be okay with numbers which are a bit off, but much cheaper?