---
title: Bookmark Computer Vision
---
[Main Page](https://hackmd.io/@hpcfung/Main_Page)
# Bookmark Computer Vision

Follow-up to [bookmark editing](https://hackmd.io/@hpcfung/bookmark_editing)

pdf to jpg
- python library: too slow (1-3 hr)
- screenshot adobe?
- in the end, cloud, free services in minutes; lower res tho (cannot customize res)?

actually lower res, CV runs faster?
however, leads to errors: eg 2.1 becomes 2.1.
also missed first 1.1

embarassingly parallel
pool: easy to parallelize

Win 10, C drive: 2-3 mins?

Win 11+D drive: 1 min?

different document tho
(eg different resolution/page size?)

parallel:
only ~3GHz tho? not max = 4.7 or so?
100% utilization

C drive or D driver faster?

manual feature engineering
library = standard tools to choose from
each book: need to come up with different strategy

eg 1.2.3 section labelled
eg section name all capital

https://www.ilovepdf.com/pdf_to_jpg

not necessary: manually allocate pages to each core? (docs)

eg class return sentence line

clash words: figure or Figure or section (rare enough, can fix manually?)

if eg Al.1, l, cannot detect

avoid too specific (stringent): may miss other cases
best resolved manually

competing
how much automation, how much manual
automation horizon in action

stress on eye

649 PAGES

dep on subtitle density

first run around 30 mins

GET FONt style, font size?

library: detect if (mostly) all caps (section title usually are)

Idea: semi-automated?
preview of page
use box: OCR that box
add bookmark

so: detection is done by human