docs: Binary image segmentation example

[iwoplaza]

• Find a no-attribution model for binary (bg/fg) segmentation that is lightweight and can be run in real-time
• Load the weights into an example, and write a compute shader to infer this network
• It will most likely be a network with convolutional layers, you can peek at how it's implemented here: https://github.com/iwoplaza/phoure
• Try to use fma 👱🤖
• The result should be a black/white mask that can be used to apply effects

[aleksanderkatan]

Here is a model that we could use, however it is quite complicated
https://aihub.qualcomm.com/models/sinet
https://github.com/cpuimage/SINet/blob/main/model.py#L4
(MIT license)

[aleksanderkatan]

Just so it does not get lost:

• ort.env.logLevel = 'verbose'; makes ONNX log the WGSL code it generates, so we could potentially use it to create our own TF/ONNX-like.
• https://netron.app/ is a nice app for viewing the layers of an ONNX model.
• https://github.com/microsoft/onnxruntime/blob/main/js/web/docs/webgpu-operators.md here is the list of what is currently implemented in WebGPU backend of ONNX.

That being said, I'm still working on the PR that uses an ONNX model with onnx-runtime.