For real, this is straight up dystopian

(Violation of)

It’s right there, clear as mud:

from wifi_densepose import WiFiDensePose

    # Initialize with default configuration
    system = WiFiDensePose()

    # Start pose estimation
    system.start()

    # Get latest pose data
    poses = system.get_latest_poses()
    print(f"Detected {len(poses)} persons")

    # Stop the system
    system.stop()

AI solves the Emporer’s Nose problem: you have no data whatsoever going in and you estimate the result!

After a bit more browsing, I found:

    # Hardware Settings
    WIFI_INTERFACE=wlan0
    CSI_BUFFER_SIZE=1000
    HARDWARE_POLLING_INTERVAL=0.1

So maybe it uses one WiFi interface to collect CSI from multiple BSSIDs? Does 802.11 support this well? (I assume you can get one-way CSI data, single-in-multiple-out, from a beacon if you really want to. [1]) Does commodity hardware support this? Do the drivers support this?

But I’d be rather impressed if that’s all that’s needed to get poses without any calibration for the actual positions of all involved devices especially if the CSI available is all of this form. This whole repo smells a bit like it’s almost 100% vibes and no content.

Wasn’t 802.11bf supposed to make real channel state information available for vendor-neutral use? What happened to it?

[1] Yes, I know, reciprocity. One-way and two-way data ought to be the same. But those nice access points almost all have at least two transmit/receive chains these days, possibly more, and they support multiple frequencies, and unless you can convince them to cooperate with you by sending known test patterns that let you disambiguate between the two antennas or at least collect vector or matrix data in a consistent basis, you don’t get to take advantage of it, and as far as I know Wi-Fi beacons don’t do this. APs do try to get something like this data for downlink MU-MIMO purposes, and stations that are receiving data with a multiple stream code get vector data fairly directly, but I’m not sure any of this works without being associated. I do wonder whether appropriate hardware can passively listen to a transmission intended for someone else and decode enough of it to extract the full CSI matrix from the transmitter to yourself.

From the author’s GitHub bio:

‘One of the most captivating aspects of AI models like GPT-4 is their ability to "hallucinate" – generating completely new ideas and concepts that go beyond mere data processing. This capability underscores AI's potential to create, not just analyze.’

‘My projects represent this space, a space of infinite possibilities only one step removed from reality.’

Rather honest I suppose.

Overnight the project went from 1.3k stars (I believe I watched it cross to this from 1.2k stars while browsing the repo last night) to 2.5k stars. It's hard for me to imagine anything besides these stars being bought.

And the green checkmarks

Who let the arms races reign?

All signs point to someone letting an LLM run wild.

Would be interested to read your findings, please do a follow-up!