| 1 #################################################################### |
|
| 2 # Predictive online PDPS for optical flow with known velocity field |
|
| 3 #################################################################### |
|
| 4 |
|
| 5 __precompile__() |
|
| 6 |
|
| 7 module AlgorithmPrimalOnly |
|
| 8 |
|
| 9 identifier = "pdps_known_primalonly" |
|
| 10 |
|
| 11 using Printf |
|
| 12 |
|
| 13 using AlgTools.Util |
|
| 14 import AlgTools.Iterate |
|
| 15 using ImageTools.Gradient |
|
| 16 |
|
| 17 using ..OpticalFlow: ImageSize, |
|
| 18 Image, |
|
| 19 pdflow! |
|
| 20 |
|
| 21 ######################### |
|
| 22 # Iterate initialisation |
|
| 23 ######################### |
|
| 24 |
|
| 25 function init_rest(x::Image) |
|
| 26 imdim=size(x) |
|
| 27 |
|
| 28 y = zeros(2, imdim...) |
|
| 29 Δx = copy(x) |
|
| 30 Δy = copy(y) |
|
| 31 x̄ = copy(x) |
|
| 32 |
|
| 33 return x, y, Δx, Δy, x̄ |
|
| 34 end |
|
| 35 |
|
| 36 function init_iterates(xinit::Image) |
|
| 37 return init_rest(copy(xinit)) |
|
| 38 end |
|
| 39 |
|
| 40 function init_iterates(dim::ImageSize) |
|
| 41 return init_rest(zeros(dim...)) |
|
| 42 end |
|
| 43 |
|
| 44 ############ |
|
| 45 # Algorithm |
|
| 46 ############ |
|
| 47 |
|
| 48 function solve( :: Type{DisplacementT}; |
|
| 49 dim :: ImageSize, |
|
| 50 iterate = AlgTools.simple_iterate, |
|
| 51 params::NamedTuple) where DisplacementT |
|
| 52 |
|
| 53 ################################ |
|
| 54 # Extract and set up parameters |
|
| 55 ################################ |
|
| 56 |
|
| 57 α, ρ = params.α, params.ρ |
|
| 58 R_K² = ∇₂_norm₂₂_est² |
|
| 59 γ = 1.0 |
|
| 60 Λ = params.Λ |
|
| 61 τ₀, σ₀ = params.τ₀, params.σ₀ |
|
| 62 |
|
| 63 τ = τ₀/γ |
|
| 64 @assert(1+γ*τ ≥ Λ) |
|
| 65 σ = σ₀*1/(τ*R_K²) |
|
| 66 |
|
| 67 println("Step length parameters: τ=$(τ), σ=$(σ)") |
|
| 68 |
|
| 69 ###################### |
|
| 70 # Initialise iterates |
|
| 71 ###################### |
|
| 72 |
|
| 73 x, y, Δx, Δy, x̄ = init_iterates(dim) |
|
| 74 init_data = (params.init == :data) |
|
| 75 |
|
| 76 #################### |
|
| 77 # Run the algorithm |
|
| 78 #################### |
|
| 79 |
|
| 80 v = iterate(params) do verbose :: Function, |
|
| 81 b :: Image, |
|
| 82 v_known :: DisplacementT, |
|
| 83 🚫unused_b_next :: Image |
|
| 84 |
|
| 85 ################## |
|
| 86 # Prediction step |
|
| 87 ################## |
|
| 88 if init_data |
|
| 89 x .= b |
|
| 90 init_data = false |
|
| 91 end |
|
| 92 |
|
| 93 pdflow!(x, Δx, y, Δy, v_known, false) |
|
| 94 |
|
| 95 ############ |
|
| 96 # PDPS step |
|
| 97 ############ |
|
| 98 |
|
| 99 ∇₂ᵀ!(Δx, y) # primal step: |
|
| 100 @. x̄ = x # | save old x for over-relax |
|
| 101 @. x = (x-τ*(Δx-b))/(1+τ) # | prox |
|
| 102 @. x̄ = 2x - x̄ # over-relax |
|
| 103 ∇₂!(Δy, x̄) # dual step: y |
|
| 104 @. y = (y + σ*Δy)/(1 + σ*ρ/α) # | |
|
| 105 proj_norm₂₁ball!(y, α) # | prox |
|
| 106 |
|
| 107 ################################ |
|
| 108 # Give function value if needed |
|
| 109 ################################ |
|
| 110 v = verbose() do |
|
| 111 ∇₂!(Δy, x) |
|
| 112 value = norm₂²(b-x)/2 + params.α*γnorm₂₁(Δy, params.ρ) |
|
| 113 value, x, [NaN, NaN], nothing |
|
| 114 end |
|
| 115 |
|
| 116 v |
|
| 117 end |
|
| 118 |
|
| 119 return x, y, v |
|
| 120 end |
|
| 121 |
|
| 122 end # Module |
|
| 123 |
|
| 124 |
|