Image samples and derivatives for direct/photometric image alignment

Specific Domains Signal and Image Processing
7

The computation itself is cheap enough that I’d expect it to be dominated by memory bandwidth if you calculate both the value and gradient at the same time

julia> function gradient_only(itp_img, locs)
           s = itp_img(locs[1]...)*0
           @inbounds for loc in locs
               s += sum(Interpolations.gradient(itp_img, loc...))
           end
           return s
       end
gradient_only (generic function with 1 method)

julia> function gradient_value(itp_img, locs)
           s = itp_img(locs[1]...)*0
           @inbounds for loc in locs
               s += itp_img(loc...)
               s += sum(Interpolations.gradient(itp_img, loc...))
           end
           return s
       end
gradient_value (generic function with 1 method)

julia> img = rand(RGB{N0f8}, 1024, 1024);

julia> locs = collect(zip(rand(axes(img, 1), 2048), rand(axes(img, 2), 2048)));

julia> itp_img = linear_interpolation(axes(img), img);

julia> @btime gradient_only($itp_img, $locs)
  69.400 μs (0 allocations: 0 bytes)
RGB{Float64}(-1.1529411764705895,-18.10980392156866,-1.8392156862744766)

julia> @btime gradient_value($itp_img, $locs)
  65.600 μs (0 allocations: 0 bytes)
RGB{Float64}(1030.321568627451,1024.2039215686275,1020.3254901960775)

Strange - calculating the value and gradient was slightly faster than calculating the gradient alone! Probably a benchmarking artifact.