csertegt3/ransac.jl

Efficient RANSAC in Julia

ransac.jl's Issues

Rework logging system

Would be greate to have loggers that can log messages between two levels (and not only above a threshold).
The used levels should be under Debug, so you can't turn them on with Debug level.
A new type of AbstractLogger, that has a method definition for shouldlog?

Add json export to example

Forgot to do that, should be easy.

Octree and point cloud should be merged in one type

Probably the octree and the point cloud (PointCloud type) should live together, for example in a RansacCloud.
Edit: they should be and can be constructed at the same time. It would eliminate some confusing parts, for example why is the levelscore of a point cloud initialized to [0.0]?

empty array can't be exported to JSON

julia> using RANSAC

julia> exportJSON(stdout, [])
ERROR: MethodError: no method matching toDict(::Array{Any,1})
Closest candidates are:
  toDict(::FittedShape) at C:\Users\Laci\.julia\dev\RANSAC\src\json.jl:10
  toDict(::ShapeCandidate) at C:\Users\Laci\.julia\dev\RANSAC\src\json.jl:18
  toDict(::ScoredShape) at C:\Users\Laci\.julia\dev\RANSAC\src\json.jl:20
  ...
Stacktrace:
 [1] exportJSON(::Base.TTY, ::Array{Any,1}) at C:\Users\Laci\.julia\dev\RANSAC\src\json.jl:58
 [2] top-level scope at REPL[2]:1

Preparing for docker - config file

I'm thinking on being able to use the package in a docker container. I'm planning something like:

1. input: a config file, and the input mesh
2. parse config file and set up the problem
3. run the iteration
4. export result to json

I have 3 options in mind for the config file format:

• ConfParser.jl - https://github.com/JuliaIO/ConfParser.jl
• TOML.jl - https://github.com/JuliaLang/TOML.jl
• YAML.jl - https://github.com/BioJulia/YAML.jl
• JSON.jl - https://github.com/JuliaIO/JSON.jl
• Pkg.TOML - TOML parser in Pkg

Either way, a RANSACParameters struct must be constructed. I'm thinking on a constructor that takes a Dict{String,Any} dictionary and uses the get(collection, key, default) method to lookup the fieldnames of the struct. Different keys (like setallepsilons) can be hardcoded this way as well. I also think, that all of this must not be that performant, because will happen only while setting up the algorithm.

With the above in mind:

• Would be great not having a ton of packages as dependency and we already have JSON.jl
• I don't like ConfParser-s output (dictionary, where the value is always an array)
• I don't know TOML.jl-s future, as it will be an stdlib package
• Pkg.TOML is not a public API, and I must find something that is supported through all the 1.x lifecycle (eg. it must work on 1.0.x, 1.x and nightly as well)
• I don't like the feel of editing a json file
• I like YAML.jl-s API (it outputs a simple dictionary)

Edit:
from the docs:

Inside f, kwargs will be a key-value iterator over a named tuple. Named tuples (as well as dictionaries with keys of Symbol) can be passed as keyword arguments using a semicolon in a call, e.g. f(x, z=1; kwargs...).

Failure due to negative sqrt

I somehow get a negative sqrt at:

function hypergeomdev(N, x, n)
    sq = sqrt((x*n*(N-x)*(N-n))/(N-1))
    (min=(x*n+sq)/N, max=(x*n-sq)/N)
end

My stack trace:

ERROR: LoadError: DomainError with -1.381857106372905e12:
sqrt will only return a complex result if called with a complex argument. Try sqrt(Complex(x)).
Stacktrace:
 [1] throw_complex_domainerror(::Symbol, ::Float64) at ./math.jl:33
 [2] sqrt at ./math.jl:573 [inlined]
 [3] hypergeomdev(::Int64, ::Int64, ::Int64) at /home/ole/.julia/packages/RANSAC/wkHhD/src/confidenceintervals.jl:54
 [4] estimatescore at /home/ole/.julia/packages/RANSAC/wkHhD/src/confidenceintervals.jl:69 [inlined]
 [5] scorecandidate(::RANSACCloud{Array{SArray{Tuple{3},Float32,1,3},1},Array{SArray{Tuple{3},Float32,1,3},1},Array{Float32,1}}, ::FittedPlane{SArray{Tuple{3},Float32,1,3},SArray{Tuple{3},Float32,1,3}}, ::Int64, ::NamedTuple{(:iteration, :common, :plane, :cone, :cylinder, :sphere),Tuple{NamedTuple{(:drawN, :minsubsetN, :prob_det, :shape_types, :τ, :itermax, :extract_s, :terminate_s),Tuple{Int64,Int64,Float64,Array{UnionAll,1},Int64,Int64,Symbol,Symbol}},NamedTuple{(:collin_threshold, :parallelthrdeg),Tuple{Float64,Float64}},NamedTuple{(:ϵ, :α),Tuple{Float64,Float64}},NamedTuple{(:ϵ, :α, :minconeopang),Tuple{Float64,Float64,Float64}},NamedTuple{(:ϵ, :α),Tuple{Float64,Float64}},NamedTuple{(:ϵ, :α, :sphere_par),Tuple{Float64,Float64,Float64}}}}) at /home/ole/.julia/packages/RANSAC/wkHhD/src/shapes/plane.jl:69

It works for some parts of my points and normals but haven't been able to figure out why it fails for the whole point cloud.

Do you have a test fail that I can check to see whether we have the same data structure? Or do you have any ideas why this might happen?

CI takes too long for PR-s

PRs and corresponding branches also run on AppVeyor and Travis, which is completely unnecessary. In theory the settings are correct, but they don't do what I want. Maybe CI behaviour should be controlled from the CI scripts.

Default values like in CGAL

CGAL has consistent default values. Maybe I should copy them.
In this file.

Struct-of-arrays vs array-of-structs

Struct-of-arrays would maybe a better way to store ScoredShapes.
There's a huge garbage collection in the findhighestscore() function, due to calling getproperty() many times.
One problem with SoA approach is, that one of the arrays would be FittedShape[], which is an abstract container.

Edit: should I use a package, or implement it myself?

Add back sampling and examples for functional validation

One day the sampling and example point clouds from RANSACBenchmark.jl should be merged back. One of the example/benchmark point clouds should be used to validate the functionality of the package by running the algorithm on it as part of CI.

Estimatescore

The score estimatation may not be correct, because the whole size of a subset is used, not just the enabled ones. Must check the paper what to use.

Overall better sampling

Sampling is currently requires time and ugly. I should completely rethink it.

• should be move to a function, in theory it could be computed from a PointCloud and the parameters.

Constructing the default `RANSACParameters` errors

julia> using RANSAC

julia> RANSACParameters()
ERROR: MethodError: no method matching RANSACParameters(::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Int64, ::Int64, ::Float64, ::Int64, ::Int64, ::Int64, ::Int64, ::Float64, ::Int64, ::Float64, ::Float64, ::Symbol, ::Int64, ::Int64, ::Float64, ::Float64, ::Bool, ::Symbol, ::Float64, ::Float64, ::Float64, ::Bool, ::Array{Symbol,1}, ::Symbol, ::Symbol)
Closest candidates are:
  RANSACParameters(::R, ::R, ::R, ::R, ::R, ::R, ::R, ::R, ::R, ::R, ::R, ::R, ::R, ::R, ::Any, ::Any, ::R, ::Any, ::Any, ::Any, ::R, ::R, ::R, ::R, ::R, ::Any, ::Any, ::Any, ::R, ::R, ::Any, ::Any, ::R, ::R, ::R, ::Any, ::Any, ::Any, ::Any) where R<:Real

Docs of Parameters.jl suggests that it should work.

Performance meta

Performance related issues. If most of it is solved, the package is ready for announcement.

• #19 - implemented in #22
• #20 - partially solved by #24
• #13 - I think I've done what I wanted in #18
• maybe #16 - implemented in ed02428, though its performance may be worse
• #18 - optimizations: merged
• 7609eeb - optimizing the cell selection
• #26 #27 - not sure, that it will help performance, but at least it will be consistent

Add more views()

There are places, where @view could be used. Should check the whole core for such parts.

Set float types

Currently I use Float64 everywhere. Maybe speed is not comparable with the paper because "everyone" uses Float32 in reverse engineering(?). Types should be propagated and Float64 shouldn't be used in code (1/2 or 1//2 instead of 0.5). A function is needed that converts the parameters to the requested type. Should it be done automatically?

change reset_rand keyword to Union{Nothing, Int}

Instead of:

function ransac(...., reset_rand=false)
    reset_rand && Random.seed!(1234)
    ...
end

do this:

function ransac(..., reset_rand::Union{Nothing,Int}=nothing)
    reset_rand === nothing || Random.seed!(reset_rand)
    ...
end

Consider using StructArrays.jl

repo link

Use of ConstructiveGeometry.jl

The ConstructiveGeometry.jl package could be used for visualization by implementing "conversion" in RANSACVisualizer.jl. It's a question if I could use it for other purposeses as well.

Construct parameters struct with degrees instead of radians

In most cases radian is used, but would be nice, if there would be a constructor, that converts from degrees to radian, so one must not do this by hand.

Use ProgressMeter.jl instead of logging

Package link.

I should check the performance, but it shouldn't really be affected.

CI rewrite

I think I use outdated CI services. Probably everything should be moved to GH actions. That maybe solves #23 without taking care of it really.

The `isntequal()` function is not working 100% as I want

From #12:
The isntequal() function is not working 100% as I want:

julia> t1 = (a1 = [1,2,3], b=4)
(a1 = [1, 2, 3], b = 4)

julia> t2 = (a1 = [2,3,1], b=4)
(a1 = [2, 3, 1], b = 4)

julia> isntequal(t1,t2)
false

The two arrays are different, but they have the same elements, therefore in this case they are equal. One workaround would be to sort the arrays:

function isntequal(t1::T1,t2::T2) where {T1<:AbstractArray, T2<:AbstractArray}
    return sort(t1) == sort(t2)
end

but sort() is not define for FittedShape types, so it's not a real workaround.

It only affects tests, so I'll just open an issue for the record.

Originally posted by @cserteGT3 in #12 (comment)

Another RANSAC implementation in Julia

This is not an issue, I just wanted to let you know that there is another RANSAC implementation in Julia: https://github.com/peterkovesi/ImageProjectiveGeometry.jl/blob/master/doc/ransac.md. It seems that they have line and plane fitting only.

How to add new primitives?

What needs to be done, to add a new shape primitive? (Just gathering the tasks, syntax and names are subject to change.)

• struct NewShape <: FittedShape: a new type
• fitshape(::Type{NewShape},...): a function that fits the primitive
• compatibles(shape::NewShape): compute the compatible points
• scorecandidate() - is it even necessary?
• refitshape()
• a way to add the new primitive to the list of primitives, that are fitted in an iteration (i.e. to RANSACParameters)
• parameters for the new shape !!! - this is the major issue