Spatial calibration · service v1
Two cameras can see the same person and disagree about where they are. I invented the first version of the spatial calibration service to bring those views into one shared frame of reference.
Open still image
The question
A person appears in a different position in each camera’s image. I wanted those observations to refer to the same place. That meant moving from image pixels to shared geometry and, ultimately, coordinates in the site.
My contribution was the first service implementation connecting camera images to a common spatial reference. During its development, the six-view experiment let me inspect the original images alongside their combined 3D representation.
The research still shows the camera images beside a fused point cloud, a set of points describing the scene in three dimensions. Coloured camera markers and positions make the relationship between the views and the reconstruction visible.
That was the useful shift: I could inspect the observations together in space, alongside the images they came from.
How it fits together
Start with views of a place. Each image describes the scene from its own viewpoint.
Bring the views into a common geometric frame so they can describe the same space.
Connect an observation in an image to a position in that shared space.
Watch Generator
Related work · Oosto era
In Generator, I explored the complementary problem: creating controlled synthetic scenes with known camera geometry and scene information. That gives perception experiments a reference to test against.
I shared the v1.0 demonstration in January 2024. Synthetic scenes and spatial calibration approach the same question from two directions: building a known world, and relating camera observations to a physical one.
Watch Generator v1.0 on YouTubeWhat this work established
The research image and public Generator film show two sides of the work: relating real camera views, and creating a world whose geometry is known. Together they show why a shared picture is only part of the problem; checking it against an independent reference matters just as much.
Evidence scope: research imagery from v1 development, not a field benchmark.