The idea: "A 20 second video of a cyclist into a full on aerodynamic "Wind Tunnel" drag test".
Wind tunnel testing dictates how the cycling industry develops everything from helmets, to bikes, to clothing. Actually testing theoretical upgrades or position changes that may make them more slippery into the wind is always a guessing game unless they have access to one of the two low speed Wind Tunnels in say, all of America, costing thousands of dollars an hour.
View the 3D Model on Sketchfab
I'm Eric Semianczuk a cyclist, and a coding enthusiast, I happened upon and refined a chain of technology involving vast AI, all Open Source tools, my own software/hardware etc. to pioneer bringing this test into anyone's home, the only barrier: having a partner or friend with a standard smartphone camera.
This process gravitated a team, and soft launched a few days ago, with no advertisement, it's even attracted it's first customer and journalist inquiry.
Here's a showcase of the pipeline when we get the video, there's still some manual editing, but I'm the process of attempting to automate it even further with additional complex, local, refined AI models and other tools (timelapsed through the boring sections):
process
This is an entirely new startup, 100% boostrapped, solving a real world problem with technology that has only been commercially viable for the past year or so.
What is the technology? It involves Photogrammetry techniques "SFM" to obtain frame to frame intrinsic camera location and orientation. Utilizing NeRF AI techniques, an AI can be over fitted to the images and relative image location/orientation to learn to generate images from locations/orientations no images were taken from. Then, after a little manual touchup, the output colored cloud of points is converted into mesh using custom, extremely low level (raw Cuda programming) programs, I developed to generate watertight mesh suitable for CFD (computational fluid dynamic) testing in the industry leading software, Open Foam.
Where did this project come from? In my freetime, as a hobby, while working a typical 9-5, I'm either coding or cycling. My first major project, with thousands of users, is https://sherpa-map.com, a world cycling routing site hosted off a server in my apartment basement with features such as using custom, local, AI ensembles to classify road surface type, complete with a custom world map using the highest quality LIDAR data available (the default's currently Google's, but it's in there).
Through building this, and building many many more tools, I discovered I could create a pipeline to turn Insta 360 imagery from atop the head of a mountain biker into a 3D representation of the entire course from a single ride through:
Even using a custom UNet model which effectively masks the users head.
However, after creating many custom ThreeJS viewers, custom compression and streaming algorithms, and custom shaders, alas, I couldn't host it on a poultry 30mb up.
Here's an example using an experimental entwine format on a viewer I cloned, NOTE this hacky example is meant for desktop, controls are WASD, mouselook: https://truesegments.com/viewer/data/denmark.html
At some point, I realized the lack of business model, the sheer amount of compute power and complexity of the pipeline (fully automated though) to create these, and thought smaller.
After learning how to use industry leading CFD software, and testing vigorously, I turned my team consisting of my roommates and friend (sys admin, frontend guy, marketing guy)'s attention to this, and started building!
We all work fulltime, so, even a quick Shopify site as a starting point isn't instant, but with validation from aerodynamics, a first customer with 0 advertisement (still working on logo/marketing materials/etc.) we can't help but feel excited!