twitwi / auto-grade

Study and/or build tools for automatic correction of exams (mcq and more)

auto-grade

Study and/or build tools for automatic correction of exams (mcq and more)

List of projects

AMC

• http://auto-multiple-choice.net/
• http://project.auto-multiple-choice.net/projects/auto-multiple-choice/
• https://bitbucket.org/auto-multiple-choice/auto-multiple-choice

Much used, active, in perl+c+glade.

Features:

• from the authoring to the grading,
• latex or simplified source,
• randomization of blocks/questions/answers,
• generate different sheets,
• rule for grading,
• checking and correcting the automatic box-checked process,
• produce reference correction,
• produce corrected individual sheets

Pain points:

• editing (not any latex), sometimes
• getting the same number of pages for each sheet (at least a warning would be good)
• making it compact
• not 100% trust in the interface etc
• cannot install cleanly 2 versions
• adding out-of-amc exercices need some hack (fake question)

Process: general

• data in .sqlite files (mostly one per step)
• layout.sqlite contains the position of most elements (boxes, corners, name field, ...)

Process: generating pdf and finding box positions

• gui entry is calcule_mep that calls "auto-multiple-choice","meptex" (which runs AMC-meptex.pl) and after detecte_mep
• it actually relies on latex generating a DOC-calage.xy file (which is kind of easy to parse)
• the name of this file is decided when running AMC-prepare.pl
• """unit is sp, such that 65536 × 72.27 sp is one inch).""" (in the doc of the automultiplechoice latex package, doc/sty/automultiplechoice.pdf)
• data access in layout.pm
  • e.g. table layout_box, column xmin,xmax,ymin,ymax
  • saved to layout.sqlite
  • it checks for instance that marks are at the same place in all pages

Things about latex: how it generates the xy file

• using /usr/share/texmf/tex/latex/AMC/automultiplechoice.sty
• wrongchoice{inanswer}{content} # inanswer = what to show in answer box for separateanswersheet
• … AMC@box{#1}{}
• … AMC@box{inanswer}{}
• …… AMC@answerBox@{}{#2}{1}{case:\AMCid@name:\the\AMCid@quest,\the\AMCrep@count}
• …… AMC@answerBox@{}{}{1}{case:\AMCid@name:\the\AMCid@quest,\the\AMCrep@count}
• ……… draw the box, and,
• ……… build a commandname based on the \AMC@shapename (can plug new ones...)
• ……… compute Z = AMCchoiceLabelFormat{#1} = AMCchoiceLabelFormat{}
• ……… AMC@shape@square{Z}{#2}{#3}{#4}
• ……… AMC@shape@square{Z}{}{1}{case:...}

The scanned box content now ends up in the db, in capture.sqlite.

How does it get the blue and pink boxes.

In AMC-detect.cc, "pink", due to illustr_mode==ILLUSTR_BOX, itself due to the main() (or the pixel version with the -k option). The AMC-detect is the default called when Subprocess.pm is used with overriding "mode", so in AMC-analyse.pl:402. Then the process (pl) gets $process->command that is parsed in .cc (via IPC) sscanf(commande,

In the end the only real solution seem to need to patch AMC-detect.cc, so we can set the default to not drawing anything (e.g., illustr_mode = 0), and after a backup of AMC-detect, do

diff AMC-detect{-ori,}.cc
77a78
> #define ILLUSTR_DEFAULT 0
800a802
>     if (illustr_mode == ILLUSTR_DEFAULT) delta = 0;
1013c1015
<   int illustr_mode=ILLUSTR_BOX;
---
>   int illustr_mode=ILLUSTR_DEFAULT;

make AMC-detect && sudo cp AMC-detect /usr/lib/AMC/exec/AMC-detect

now consolidated in a custom Dockerfile

Alternative, prototypes etc

---

In Python, single file using opencv:

https://github.com/RanaSamy/Multiple-choice-auto-correction

• uses cv2.HoughCircles
• get the extreme coordinates of these
• rectify with cv2.getRotationMatrix2D and cv2.warpAffine
• crop, blur
• uses 3 different crops (not sur what regions they are)
• then cv2.adaptiveThreshold
• then it seems a little ad'hoc

---

Helping building AMC questionaires

python3

https://github.com/jarthurgross/amc_question_creator

---

(web)

https://github.com/avitheque/amc-builder

https://amc-builder.avitheque.net/