Mani-RRI / board-to-latex

Turn pictures of math into LaTeX

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

board-to-latex

Take a picture of a blackboard or whiteboard, and convert it into a LaTeX document!

Hack the North 2019 hackathon project. As such, please excuse the hackathon-level code quality.

Usage

NOTE: Due to Google API changes, this project is no longer working, and development is paused indefinitely

Terminal UI:

./app/run_tui.sh <image_file>

For the web UI, see the instructions in the app folder.

Samples

Input image:

Input image

Output pdf:

Output

Output latex:

\begin{thm}[Rank-Nullity]
Let v be a finite dimensional vector space over Ft. Let $T : V \rightarrow W$ be linear transformation. Then: 
$$\operatorname { rank } ( T ) + \operatorname { nullity } ( T ) = \operatorname { dim } ( V )$$

\begin{proof}
Exercise for the reader. 
$$\sum _ { n = 0 } ^ { \infty } n = \frac { - 1 } { 12 }$$

\end{proof}
\end{thm}

How it works

Due to hackathon time constraints, we were not able to train a neural network to accomplish this goal end-to-end as desired. Instead, we aimed for usability.

Current OCR implementations at the time, like MathPix and Google OCR, did not provide satisfactory results:

  • Mathpix was good at math, but at the time had trouble with layout and general text
  • Google OCR was good at general text but could not do math

We combined the two, and developed a method that allows the app to split images into math components and text components to send off to either Mathpix/Google respectively. Each service would receive images to recognize that they are good at, and then our app would combine them after the fact. The end results, as seen in the samples section, worked quite nicely - it was a one-stop solution unlike mathpix at the time, and we put it into good use for several math lectures after the hackathon up until the API breakage.

About

Turn pictures of math into LaTeX

License:GNU Affero General Public License v3.0


Languages

Language:Python 60.9%Language:JavaScript 35.5%Language:HTML 2.9%Language:Shell 0.7%