Azure Kinect DK

~ the obituary ~

this one lands closer to my desk than most graves on this site. i work on computer vision for edge devices, which means depth sensors are a tool i think about rather than a gadget i read about. azure kinect dk was the cheap end of that tool category for a few years, and then it was not for sale.

microsoft announced it at mobile world congress in february 2019 and opened preorders the same week. it reached buyers in the united states in march 2020 at three hundred and ninety nine dollars, then the united kingdom, germany, japan and china. the shape was a small aluminum slab with a usb-c cable, nothing like the long black bar that sat under a television in 2011.

inside were a one megapixel time-of-flight depth camera, a twelve megapixel rgb camera, a seven microphone circular array and an inertial measurement unit. the depth sensor was the part that mattered. it used the indirect time-of-flight approach microsoft had been developing since hololens, packaged for people building robots, retail analytics, physical therapy tools, warehouse systems and academic papers.

the name was azure for a reason. the kit was sold as the front end of a cloud pipeline, and microsoft wanted the sensor feeding its cognitive services rather than sitting on a workbench. most of the people who bought one used it offline, on a workstation, with a body tracking sdk that needed an nvidia gpu to run at all. the cloud framing was marketing laid over a very good piece of hardware.

~ the case file ~

~ the third kinect ~

kinect for xbox 360 arrived in november 2010 and moved eight million units in sixty days, faster than any consumer electronics device had sold to that point. it was bought as a party toy and abandoned as one. the second version came bolted to the xbox one in 2013 at four hundred and ninety nine dollars whether you wanted it or not, and microsoft unbundled it within a year.

what kept the name alive through both disappointments was that researchers had quietly discovered the first one. a structured-light depth camera for a hundred and fifty dollars was an absurd bargain in 2010, and within months there were open drivers, robotics labs, art installations and thousands of papers built on hardware sold as an xbox accessory. azure kinect dk was microsoft finally selling to that audience on purpose. selling to it on purpose turned out to be a smaller business than selling to it by accident.

~ the actual end ~

in august 2023 microsoft told customers the azure kinect dk was being discontinued. units would be available until the end of october or until stock ran out, and spare parts would come from third parties after that. it was not a dramatic ending. the sdk stayed downloadable, and microsoft committed to maintenance updates for the sdk and the firmware through 2027, which is a longer support tail than most dead hardware gets.

the reason given was that microsoft had licensed the indirect time-of-flight sensing behind the device to hardware partners, so equivalent cameras would come from other companies. this is true and it is also an exit. microsoft stopped building the box and kept the parts of the stack it cared about, the silicon design at one end and the azure services at the other.

~ the next chapter ~

the depth camera category did not die with it. intel realsense, orbbec, luxonis and stereolabs all sell into the space azure kinect occupied, and orbbec's femto line was built on microsoft's own time-of-flight module with microsoft naming it the migration path. anyone with a working unit can still use it, and the used market has kept prices high for exactly that reason.

what did die is the specific thing microsoft was uniquely able to do, which is subsidize a research instrument by pretending it is something else. the first kinect was cheap because it was an xbox accessory carrying console economics. the third was priced like what it actually was, and priced like what it actually was, it sold like a niche developer kit. three attempts, and the only version that reached scale was the one microsoft did not build for the people who ended up using it.

@}-->--

              )
             (_)
            .-'-.
            |   |
            |   |
            |   |
            |   |
          __|   |__   .-.
       .-'  |   |  `-:   :
      :     `---'     :-'
       `-._       _.-'
           '""""""

« previous  ·  random  ·  next »