GitHub

Photometric stereo

This example illustrates photometric stereo for Lambertian surfaces using the Julia language.

This method determines the surface normals of an object from 3 or more pictures of the object taken with different lighting directions.

This demo follows the "uncalibrated" approach of Hayakawa, JOSA, 1994 that treats the lighting directions as being unknown, unlike the original least-squares approach of Woodham, 1980.

This page comes from a single Julia file: photometric3.jl.

You can access the source code for such Julia documentation using the 'Edit on GitHub' link in the top right. You can view the corresponding notebook in nbviewer here: photometric3.ipynb, or open it in binder here: photometric3.ipynb.

Setup

Add the Julia packages used in this demo. Change false to true in the following code block if you are using any of the following packages for the first time.

if false
    import Pkg
    Pkg.add([
        "Downloads"
        "InteractiveUtils"
        "LinearAlgebra"
        "LaTeXStrings"
        "MIRTjim"
        "NPZ"
        "Plots"
        "Printf"
        "Random"
    ])
end

Tell Julia to use the following packages. Run Pkg.add() in the preceding code block first, if needed.

using Downloads: download
using InteractiveUtils: versioninfo
using LaTeXStrings
using LinearAlgebra: Diagonal, svd, svdvals, rank, norm, pinv
using MIRTjim: jim, prompt
using NPZ: npzread
using Plots: gui, plot, scatter, scatter!, ylims!, cgrad, default, RGB, savefig
using Plots.PlotMeasures: px
using Printf: @sprintf
using Random: seed!
default(); default(label="", markerstrokecolor=:auto)

The following line is helpful when running this jl-file as a script; this way it will prompt user to hit a key after each image is displayed.

isinteractive() && prompt(:prompt);

Ground-truth surface normals

Load ground truth surface normal vectors of "bunny" data used in 2012 CVPR paper by Ikehata et al..

if !@isdefined(gt_normal_bunny)
    url = "https://github.com/yasumat/RobustPhotometricStereo/raw/master/data/bunny/gt_normal.npy"
    tmp = npzread(download(url))
    i1, i2 = 32:256-25, 33:212 # crop to reduce compute
    tmp = permutedims(tmp, [2, 1, 3]) # "transpose"
    gt_normal_bunny = tmp[i1, i2, :] # crop
end;

Create hemisphere to augment the bunny data.

function hemi_normal(x, y ; # surface normal of a hemi-ellipsoid
    xh = 20f0, yh = xh, zh = xh,
)
    tmp = (x/xh)^2 + (y/yh)^2
    if tmp < 1
        z = zh * sqrt(1 - tmp)
        tmp = [x/xh^2, y/yh^2, z/zh^2]
        return tmp / norm(tmp)
    end
    return [0, 0, 0]
end;

if true
    rh = 20
    x = -rh:rh
    y = x
    tmp = hemi_normal.(x, y'; xh=rh)
    tmp = reduce(hcat, tmp)
    tmp = reshape(tmp', 2rh+1, 2rh+1, 3)
end;

Define ground truth normals as combination of bunny normals and hemisphere normals.

if !@isdefined(gt_normal)
    gt_normal = copy(gt_normal_bunny)
    gt_normal[176 .+ x, 25 .+ y, :] = tmp
    nx, ny = size(gt_normal)[1:2]
    shape2 = x -> reshape(x, prod(size(x)[1:2]), :)
    shape3 = x -> reshape(x, nx, ny, :)
end;

The three images are the x, y, and z components

pn_gt = jim(gt_normal; title="Ground-truth normals", nrow=1,
    xticks = false, yticks = false, labelfontsize = 16, tickfontsize=12,
    left_margin = 20px, right_margin = 30px,
#  xaxis=L"x", yaxis=L"y",
    size=(600,200), clim=(-1,1), colorbar_ticks=-1:1,
)
# savefig(pn_gt, "photometric3_gt.pdf")
Example block output

Surface normals are meaningful only where the object is present, so we determine an object "mask".

mask = dropdims(sum(abs, gt_normal, dims=3), dims=3) .> eps(Float32)
pm = jim(mask, "Mask"; cticks=0:1)
Example block output

Verify that the surface normals are unit norm (within the mask).

@assert maximum(abs, sum(abs2, gt_normal, dims=3)[vec(mask)] .- 1) < 1e-12

View angle of surface normal w.r.t. z-axis

if true
    tmp = sqrt.(sum(abs2, gt_normal[:,:,1:2]; dims=3))
    tmp = rad2deg.(atan.(tmp, gt_normal[:,:,3]))
    pza = jim(tmp; title="Angle of surface normal w.r.t. z axis",
        ctitle="degrees", cticks=0:30:90)
end
Example block output

Lighting directions

Define lighting directions for simulated object views.

function light_vector( ;
    θ = rand() * 2π,
    r = 1/sqrt(2),
    xoffset = 0.2,
    yoffset = 0.2,
    x = r * cos(θ) + xoffset,
    y = r * sin(θ) + yoffset,
)
    z = sqrt(1 - x^2 - y^2)
    return [x, y, z]
end;

Deliberately asymmetric directions to aid testing

if !@isdefined(Ltrue)
    nlight = 12 # number of lighting directions
    Ltrue = [light_vector(; θ=il/nlight*1.7π, r=0.5-0.1*il/nlight) for il in 1:nlight]
    Ltrue = reduce(hcat, Ltrue) # (3, nlight)
    @assert maximum(abs, sum(abs2, Ltrue; dims=1) .- 1) < 9eps()
    @show extrema(Ltrue[3,:])
end;
extrema(Ltrue[3, :]) = (0.6437950455988092, 0.9894696470170959)

Plot lighting directions

tmp = range(0, 2π, 361)
plot(cos.(tmp), sin.(tmp); aspect_ratio=1, color=:black,
    xaxis = (L"x", (-1,1)),
    yaxis = (L"y", (-1,1)),
    title = "Lighting directions",
)
pl_gt = scatter!(Ltrue[1,:], Ltrue[2,:]; label = "True", color=:red)
Example block output
prompt()

Synthesize images

Synthesize image data using Lambertian reflectance model.

For the Lambertian model, each pixel value is proportional to the inner product of the lighting direction with the corresponding surface normal.

If that inner product is zero for some pixels, then the image contains "shadows" at those pixels. Wu et al., ACCV, 2011 describe the max(⋅,0) operation below as "attached shadows."

if !@isdefined(images)
    images_ideal = shape2(gt_normal) * Ltrue # hypothetical Lambertian
    images_ideal ./= maximum(images_ideal) # normalize
    svdval_ideal = svdvals(images_ideal)
    images = max.(images_ideal, 0) # "shadows" if lighting is ≥ 90° from normal
    svdval_images = svdvals(images)
    images = shape3(images)
end;

Note the different shadings in the different images. Obviously the bunny cannot "jump around" during the imaging...

pd = jim(images; title="Images for $nlight different lighting directions",
#  xticks = false, yticks = false, tickfontsize=12, # book
    caxis=("Intensity", (0,1), 0:1),
)

# savefig(pd, "photometric3_data.pdf")
Example block output

Low-rank structure

Examine the singular values. Ideally (i.e., ignoring shadows), there would be (at most) 3 nonzero singular values, because images_ideal is the product of (npixel × 3) object normal matrix with a (3 × nlight) lighting direction matrix, so its rank is at most 3.

With self-shadow effects, e.g., Barsky & Petrou, 2003, T-PAMI, there are 3 dominant singular values and additional small but non-negligible values. The images matrix rank is not exactly 3 because of the shadow effects.

if !@isdefined(ps)
    ps1 = scatter(svdval_ideal, label="Ideal", color=:blue,
       xaxis=(L"k", (1,nlight), [1,3,4,nlight]),
       yaxis=(L"σ_k",), marker=:x, widen=true,
       title="Singular values")
    scatter!(ps1, svdval_images, label="Realistic", color=:red)

    good = all(>(0), images, dims=3)
    images_good = shape2(images)[vec(good),:]
    svdval_good = svdvals(images_good)
    scatter!(ps1, svdval_good, label="Good pixels", color=:green)
    ps4 = deepcopy(ps1)
    ylims!(ps4, (0,5); title="zoom")
    ps = plot(ps1, ps4)
end
Example block output
prompt()

# savefig(ps, "photometric3_svdvals.pdf")

Rank-3 approximation

To make a low-rank approximation, collect image data into a npixel × nlight matrix and use a SVD.

Estimate the lighting directions using only the pixels with no shadows.

tmp = svd(images_good)
light1 = tmp.Vt[1:3,:] # right singular vectors
normal1 = tmp.U[:,1:3] * Diagonal(tmp.S[1:3])
@assert norm(images_good - normal1 * light1) / norm(images_good) < 9eps()

Estimate lighting directions

Apply method of Hayakawa, JOSA, 1994 to resolve non-uniqueness issue, under the simplifying assumption (satisfied here) that the light intensity is the same for all lighting directions.

That method does not explicitly exploit the fact that $A'A$ is positive semi-definite. Challenge: develop method that does use that property.

Abig = reduce(vcat, map(c -> kron(c', c'), eachcol(light1)))
tmp = Abig \ ones(nlight)
B = reshape(tmp, 3, 3) # B = A'A
@assert B ≈ B' # (symmetry check)
A = sqrt(B)
@assert A'A ≈ B
light2 = A * light1
@assert maximum(abs, sum(abs2, light2, dims=1) .- 1) < 30eps()

normal2 = normal1 * inv(A)
@assert norm(images_good - normal2 * light2) / norm(images_good) < 10eps()
@assert maximum(abs, norm.(eachrow(normal2)) .- 1) < 9e-6 # already unit norm!

As described in Hayakawa, JOSA, 1994, the estimated lighting and surface normals are in an arbitrary 3D coordinate system. To display them in a useful way, we use the Procrustes method to align the coordinate system with that of the original lighting.

if true
    tmp = Ltrue[:,1:3] * light2[:,1:3]' # use just the first 3 sources
    tmp = svd(tmp)
    tmp = tmp.U * tmp.Vt
    light3 = tmp * light2
    normal3 = normal2 * tmp'
    @assert norm(images_good - normal3 * light3) / norm(images_good) < 10eps()
end

Plot estimated lighting directions (after coordinate system alignment)

tmp = deepcopy(pl_gt)
pl = scatter!(tmp, light3[1,:], light3[2,:],
    marker = :x, label = "Estimated", color=:blue)
Example block output
prompt()

Estimate surface normals.

Having estimated the lighting directions, return to estimate the surface normals for all pixels, not just the "good" pixels.

normal3 = shape3(shape2(images) * pinv(light3));

Examine the estimated surface normals. The accuracy is very good, except in the shadow regions.

pn_hat = jim(normal3; nrow=1, title="Estimated normals",
    xticks = false, yticks = false, labelfontsize = 16, tickfontsize=12,
    left_margin = 20px, right_margin = 30px,
#  xaxis=L"x", yaxis=L"y",
    size=(600,200), clim=(-1,1), colorbar_ticks=-1:1,
)
# savefig(pn_hat, "photometric3_hat.pdf")
RGB255(args...) = RGB((args ./ 255)...)
color = cgrad([RGB255(230, 80, 65), :black, RGB255(23, 120, 232)])
pn_d = jim(normal3 - gt_normal; nrow=1, title="Difference", color,
    xticks = false, yticks = false, labelfontsize = 16, tickfontsize=12,
    left_margin = 20px, right_margin = 30px,
    xaxis=L"x", yaxis=L"y", size=(600,200), clim=(-1,1), colorbar_ticks=-1:1,
)
pn = jim(
 pn_gt,
 pn_hat,
 pn_d,
 layout=(3,1),
 size=(550, 600),
)

# savefig(pn, "photometric3_pn1.pdf")
Example block output

This demo illustrates the utility of the SVD and low-rank matrix approximation.

More advanced methods handle shadows by allowing sparse errors, e.g.,

or handle more general lighting conditions, e.g.,

Exercise

Apply the method described above to the bunny data used in Ikehata et al., CVPR 2012. This is a set of 50 images under various lighting directions.

As a starting point, here we load that data.

if !@isdefined(images_bunny)
    url0 = "https://github.com/yasumat/RobustPhotometricStereo/raw/master/data/bunny/bunny_lambert/image000.npy"

    index_bunny = 0:5:45 # just load 10 of the 50
    nlight_bunny = length(index_bunny)
    tmp = download(url0)
    x = npzread(tmp)

    dim = size(x)[1:2]
    images_bunny = zeros(Float32, dim..., nlight_bunny)
    images_bunny[:,:,1] = x[:,:,1]'

    for (iz, index) in enumerate(index_bunny[2:end])
        id3 = @sprintf("%03d", index)
        @show id3
        url1 = replace(url0, "000" => id3)
        xtmp = npzread(download(url1))
        images_bunny[:,:,iz+1] = xtmp[:,:,1]'
    end
    images_bunny ./= maximum(images_bunny) # normalize
end
pb = jim(images_bunny; title="Images for different lighting directions")
Example block output

For reference, here are the ground truth lighting directions.

if !@isdefined(gt_lights)
    url = "https://github.com/yasumat/RobustPhotometricStereo/raw/master/data/bunny/lights.npy"
    gt_lights = npzread(download(url))
    gt_lights = gt_lights'[:,index_bunny .+ 1]
end
pl_gtb = scatter(eachrow(gt_lights)...,
   xaxis = (L"x", (-0.8, 0.8), (-1:1)*0.8),
   yaxis = (L"y", (-0.8, 0.8), (-1:1)*0.8),
   zaxis = (L"z", (0.4, 1.0), [0.4, 0.69, 0.96]),
   title = "True lighting directions",
)
Example block output
prompt()

If needed, here is the url for the mask:

  • https://raw.githubusercontent.com/yasumat/RobustPhotometricStereo/master/data/bunny/mask.png

Reproducibility

This page was generated with the following version of Julia:

using InteractiveUtils: versioninfo
io = IOBuffer(); versioninfo(io); split(String(take!(io)), '\n')
11-element Vector{SubString{String}}:
 "Julia Version 1.11.1"
 "Commit 8f5b7ca12ad (2024-10-16 10:53 UTC)"
 "Build Info:"
 "  Official https://julialang.org/ release"
 "Platform Info:"
 "  OS: Linux (x86_64-linux-gnu)"
 "  CPU: 4 × AMD EPYC 7763 64-Core Processor"
 "  WORD_SIZE: 64"
 "  LLVM: libLLVM-16.0.6 (ORCJIT, znver3)"
 "Threads: 1 default, 0 interactive, 1 GC (on 4 virtual cores)"
 ""

And with the following package versions

import Pkg; Pkg.status()
Status `~/work/book-la-demo/book-la-demo/docs/Project.toml`
  [6e4b80f9] BenchmarkTools v1.5.0
  [aaaa29a8] Clustering v0.15.7
  [35d6a980] ColorSchemes v3.27.1
⌅ [3da002f7] ColorTypes v0.11.5
⌃ [c3611d14] ColorVectorSpace v0.10.0
  [717857b8] DSP v0.7.10
  [72c85766] Demos v0.1.0 `~/work/book-la-demo/book-la-demo`
  [e30172f5] Documenter v1.7.0
  [4f61f5a4] FFTViews v0.3.2
  [7a1cc6ca] FFTW v1.8.0
  [587475ba] Flux v0.14.25
  [a09fc81d] ImageCore v0.10.4
  [71a99df6] ImagePhantoms v0.8.1
  [b964fa9f] LaTeXStrings v1.4.0
  [7031d0ef] LazyGrids v1.0.0
  [599c1a8e] LinearMapsAA v0.12.0
  [98b081ad] Literate v2.20.1
  [7035ae7a] MIRT v0.18.2
  [170b2178] MIRTjim v0.25.0
  [eb30cadb] MLDatasets v0.7.18
  [efe261a4] NFFT v0.13.5
  [6ef6ca0d] NMF v1.0.3
  [15e1cf62] NPZ v0.4.3
  [0b1bfda6] OneHotArrays v0.2.5
  [429524aa] Optim v1.10.0
  [91a5bcdd] Plots v1.40.9
  [f27b6e38] Polynomials v4.0.11
  [2913bbd2] StatsBase v0.34.3
  [d6d074c3] VideoIO v1.1.0
  [b77e0a4c] InteractiveUtils v1.11.0
  [37e2e46d] LinearAlgebra v1.11.0
  [44cfe95a] Pkg v1.11.0
  [9a3f8284] Random v1.11.0
Info Packages marked with ⌃ and ⌅ have new versions available. Those with ⌃ may be upgradable, but those with ⌅ are restricted by compatibility constraints from upgrading. To see why use `status --outdated`

This page was generated using Literate.jl.