Contour segmentation is fast but not so precise
Lightweightness comes with a price. Despite contour segmentation being fast, it isn't so stable and suffers from pattern leaking and other things.
Experiment deep learning approaches with speed in mind
Maybe YOLO (focused on the bounding boxes, not the classification) or Encoder-Decoder (like U-Nets) architectures should be tested to check if it can be better on segmentation (almost certain) but also fast (uncertain), without using GPUs.
Currently the max resolution is set to FullHD (1920x1080), while corner supports lower resolutions. Video streams higher than 1920x1080 are downscaled to this resolution in real-time;
The idea here is to unlock higher resolutions and check the overall system performance;
Also the user should have the opportunity to choose the work resolution.
Today only centimeters are supported corner must be more well localized and supports other units of measure, like foot, inches, etc. It also must support more metric units subdivisions.
Create conversion methods
This is pretty easy and straightforward and corner should let the users chose their units.