Could you point me to a paper where they successfully segment random webpages in the wild semantically? Or other kinds of documents?
I expect that to be significantly different from the task of segmenting objects in a photo, because most of the boundaries are not revealed by texture differences, but implicit in the arrangement.
Most of the work is in document segmentation:
https://arxiv.org/abs/1704.01474
https://arxiv.org/abs/1711.07695
https://arxiv.org/abs/1710.07363
I'm pretty sure I've seen some webpage work too, but I don't have the reference.
Those are all about segmenting the text on a page from the background, and not semantic segmentation, which I thought you were talking about ("segment the page and decide what is the navigation").
Finding the text in a screenshot of a webpage should be easier than doing it for handwritten documents; it's the "decide what is the navigation" part that I'm doubtful about.
The exciting thing is that in this case it doesn't need to be. Segmentation (which is solved) is sufficient to identify semantically separate parts of the page, and then OCR can be used to place the text from the html on the representation of the page.