PROBLEM
The Visual segmentation does not find Sentences, that start on one page and end on the next in a PDF-based visual (and probably also in HTML-based visuals having e.g. multi-column layouts) because structurally the sentences may have other content between them (header, footer, page-numbering etc.)

SOLUTION
After the "normal" segmentation make another "brute force" run with the leftover segments that iterates to each of those and splits them trying to find both parts. Make sure there are threshholds so e.g. a failed general segmentation (e.g. wrong source/target mapping, wrong file) will not search "for hours" in this mode.

HINT
This issue could/should be solved together with https://jira.translate5.net/browse/TRANSLATE-5410