First demo of our new aerial Visual-Inertial Positioning System (VIO+VPS): 40m positioning accuracy (CEP) in a 50 kilometer flight using a camera, consumer-grade IMU and a barometer! No GPS

@907tothe703 Good question. This solution is not suitable for marine navigation, since VIO does not work at sea. Combining IMU / INS with star trackers probably works in the marine context, but it's not the same thing as this method

@BradBitler Yes, this can be integrated into ArduPilot, PX4 or other flight controller, but there are a extra hoops that you have to jump through compared to a using a GPS module. We offer commercial support for that kind of integration projects.

@NickParkerPrint This demo used a 100 x 100 km² map, which easily fits into an SD card (a 1 km x 1000 km map would take practically the same space). The ballpark size is "gigabytes" and is adjustable with a few settings that allow balancing between CPU, disk and accuracy.

Also joined Bluesky:

@jeremiahjw This is not gyroscope and accelerometer data. It's ARKit output (which is computed from that and camera data using VISLAM, and fine for most purposes). If you actually want raw IMU data + video, our recording app does that:

As shown in the video, we actually had to use two smartphones to manage the engine vibration noise: one in the cockpit and another below the plane. This was an offline demo, but a typical mobile phone processor is also powerful enough to run the method in real time

A longer version of the video is available on YouTube (corrected link):

More technical details in the next post

The software in this test consists of two main components: 1) Visual-Inertial Odometry (VIO), which is already available in our commercial SDK. 2) and a new Visual Positioning System (VPS) component, which uses the reference map. This component is not a part of our standard SDK.

@scottcfoley There is indeed a some 60's tech in the video: the airplane (built in 1967). DSMAC and modern VPS are based on the same high-level principle but are quite different under the hood.

RT @zhenjun_zhao: gsplat: An Open-Source Library for Gaussian Splatting @vickie_ye_, @ruilong_li, Justin Kerr, @_maturk, Brent Yi, @pan_zh…

@Luckyballa Check out (Python package not released yet so you'll have to compile it yourself). Based on The method allows extracting meshes in arbitrary octrees instead of dense voxel grids.

Happy to announce that Gaussian Splatting on the Move has been accepted to #ECCV2024 🎉 Project page

@oskurovic Short answer: Yes. Longer answer: For the purpose of rendering the video, the data was replayed through the SDK offline, but with settings that would also run in real-time in the Jetson/Qualcomm CPUs in those payload devices (no GPU processing used)