A pytorch / python implementation of marching squares using conv2d as the workhorse
This algorithm is useful for finding the points around the permimeter of an instance segmentation mask. It is available in scikit-image as find_contours but requires converting to a numpy array. This can be annoying if
you want to keep the data on its device.
A big part of marching squares involves iterating over pixels to find and classify boundary points. This can be achieved by convolving the following 2x2 kernel
[[1, 2],
[4, 8]]
and using the resulting values to look up relative coordinate edges.
The process of joining these edges into contours is still just a for loop, but usually there are few enough edges in the contour that this step is negligible.
- Does not apply interpolation
- Is not differentiable (could it be?)
- Cannot be compiled using dynamo or torchscript (could it be?)