add tests for chessboard corner detection

[yakir12]

This is my first PR in hopes to add a full test-suit for the opencv camera calibration functions. My secret motive is that I've found that OpenCV.calibrateCamera randomly fails (i.e. return nonsensical values), and I hope to unearth the issue together with the maintainers of this repo.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 18.75%. Comparing base (0bed41b) to head (8c4babb).

Additional details and impacted files

@@           Coverage Diff           @@
##           master      #52   +/-   ##
=======================================
  Coverage   18.75%   18.75%           
=======================================
  Files           3        3           
  Lines          32       32           
=======================================
  Hits            6        6           
  Misses         26       26           

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[yakir12]

Super weird. I noticed that just chess5.png was failing, and that it was first in the list... So I tried running the detection once without comparing its results to the saved coordinates of the corners, and now it works. It's as if there is a need for some initialization for it to work properly...?

[yakir12]

(@timholy, you asked me to ping you directly)

[timholy]

Thanks for the PR! It mostly looks good. I've left a few code comments, but one overall comment:

together with the maintainers of this repo

This repo has no maintainers. There are people like me who have merge privileges via JuliaImages, but I don't know or use OpenCV. It would be great if this package got a real maintainer.

I don't know how to fix the "must be initialized" bug, sadly.

[timholy]

Out of curiosity, could this be declared OpenCV.Mat(Matrix{Tuple{Float32,Float32}}(undef, n_corners))? And then avoid the reshaping/eachslice below?

[yakir12]

Currently, OpenCV.Mat can only accept 3D arrays, not matrices:

  MethodError: no method matching OpenCV.Mat(::Matrix{Tuple{Float32, Float32}})
  The type `OpenCV.Mat` exists, but no method is defined for this combination of argument types when trying to construct it.
  
  Closest candidates are:
    OpenCV.Mat(::AbstractArray{T, 3}) where T<:Union{Float32, Float64, Int16, Int32, Int8, UInt16, UInt8}
     @ OpenCV ~/.julia/artifacts/dcc70861ed0ffc5427898f6aeb436620f33fa02f/OpenCV/src/Mat.jl:13

So that's not an option.

[timholy]

You should be able to do

corners = Matrix{Tuple{Float32,Float32}}(undef, n_corners)
ret, cv_corners = OpenCV.findChessboardCorners(gry, cv_n_corners, OpenCV.Mat(reinterpret(reshape, Float32, corners)), 0)
@assert ret "Failed to detect any corners!"
return corners

if you think that's a little bit clearer/cleaner. It certainly makes inference's job easier.

[yakir12]

I agree and understand. I'm not even sure there's any need to know much about OpenCV to fix this. As I've noticed elsewhere, opencv via PythonCall works just fine, so it seems like something in the porting of opencv to Julia went a little off. My goal is to add a bunch of tests to this module that might highlight and insulate the issues in hand so that it'll be easier for anyone to resolve the issue. The main one being this row.