Just back from vacation, sorry for the delay in my response.
It works fine on Windows, which configures the camera as 1:1 ratio.
I am only seeing the issue when deploying to a Linux (Ubuntu 20.04.4 LTS) computer.
It this case, the camera is configured as 640:480 ratio, and the detection results are displayed on a 480:480 grid within this image.
Hope this helps
This definitely looks like a bug regarding the squashing to square for training/inference, but then not doing the equivalent projection for inference in browser… I’d bet if we scaled those horizontal values by 640/480 it’d line up…
@louis ; does the face detection project you mention work with a square input for in browser inference?