Mon, 22 Apr 2024 13:28:58 +0300
stable interval
0 | 1 | ###################################################################### |
2 | # Predictive online PDPS for optical flow with unknown velocity field | |
3 | ###################################################################### | |
4 | ||
5 | __precompile__() | |
6 | ||
7 | module AlgorithmBothGreedyV | |
8 | ||
9 | identifier = "pdps_unknown_greedyv" | |
10 | ||
11 | using Printf | |
12 | ||
13 | using AlgTools.Util | |
14 | import AlgTools.Iterate | |
15 | using ImageTools.Gradient | |
16 | ||
17 | using ..OpticalFlow: Image, | |
18 | ImageSize, | |
19 | DisplacementConstant, | |
20 | DisplacementFull, | |
21 | pdflow!, | |
22 | horn_schunck_reg_prox!, | |
23 | pointwise_gradiprod_2d!, | |
24 | filter_hs | |
25 | ||
26 | using ..Algorithm: step_lengths | |
27 | ||
28 | ######################### | |
29 | # Iterate initialisation | |
30 | ######################### | |
31 | ||
32 | function init_displ(xinit::Image, ::Type{DisplacementConstant}) | |
33 | return xinit, zeros(2) | |
34 | end | |
35 | ||
36 | function init_displ(xinit::Image, ::Type{DisplacementFull}) | |
37 | return xinit, zeros(2, size(xinit)...) | |
38 | end | |
39 | ||
40 | function init_rest(x::Image, u::DisplacementT) where DisplacementT | |
41 | imdim=size(x) | |
42 | ||
43 | y = zeros(2, imdim...) | |
44 | Δx = copy(x) | |
45 | Δy = copy(y) | |
46 | x̄ = copy(x) | |
47 | ||
48 | return x, y, Δx, Δy, x̄, u | |
49 | end | |
50 | ||
51 | function init_iterates( :: Type{DisplacementT}, xinit::Image) where DisplacementT | |
52 | return init_rest(init_displ(copy(xinit), DisplacementT)...) | |
53 | end | |
54 | ||
55 | function init_iterates( :: Type{DisplacementT}, dim::ImageSize) where DisplacementT | |
56 | return init_rest(init_displ(zeros(dim...), DisplacementT)...) | |
57 | end | |
58 | ||
59 | ############ | |
60 | # Algorithm | |
61 | ############ | |
62 | ||
63 | function solve( :: Type{DisplacementT}; | |
64 | dim :: ImageSize, | |
65 | iterate = AlgTools.simple_iterate, | |
66 | params::NamedTuple) where DisplacementT | |
67 | ||
68 | ###################### | |
69 | # Initialise iterates | |
70 | ###################### | |
71 | ||
72 | x, y, Δx, Δy, x̄, u = init_iterates(DisplacementT, dim) | |
73 | init_data = (params.init == :data) | |
74 | ||
75 | # … for tracking cumulative movement | |
76 | if DisplacementT == DisplacementConstant | |
77 | ucumul = [0.0, 0.0] | |
78 | else | |
79 | ucumul = [NaN, NaN] | |
80 | end | |
81 | ||
82 | ############################################# | |
83 | # Extract parameters and set up step lengths | |
84 | ############################################# | |
85 | ||
86 | α, ρ, λ, θ, T = params.α, params.ρ, params.λ, params.θ, params.timestep | |
87 | R_K² = ∇₂_norm₂₂_est² | |
88 | γ = 1 | |
89 | τ, σ, σ̃, ρ̃ = step_lengths(params, γ, R_K²) | |
90 | ||
91 | kernel = params.kernel | |
92 | ||
93 | #################### | |
94 | # Run the algorithm | |
95 | #################### | |
96 | ||
97 | b_next_filt=nothing | |
98 | ||
99 | v = iterate(params) do verbose :: Function, | |
100 | b :: Image, | |
101 | 🚫unused_v_known :: DisplacementT, | |
102 | b_next :: Image | |
103 | ||
104 | #################################### | |
105 | # Smooth data for Horn–Schunck term | |
106 | #################################### | |
107 | ||
108 | b_filt, b_next_filt = filter_hs(b, b_next, b_next_filt, kernel) | |
109 | ||
110 | ################## | |
111 | # Prediction step | |
112 | ################## | |
113 | ||
114 | if init_data | |
115 | x .= b | |
116 | init_data = false | |
117 | end | |
118 | ||
119 | pdflow!(x, Δx, y, Δy, u, params.dual_flow) | |
120 | ||
121 | # Predict zero displacement | |
122 | u .= 0 | |
123 | if params.prox_predict | |
124 | ∇₂!(y, x) | |
125 | @. y = (y + σ̃*Δy)/(1 + σ̃*(ρ̃+ρ/α)) | |
126 | proj_norm₂₁ball!(y, α) | |
127 | end | |
128 | ||
129 | ############ | |
130 | # PDPS step | |
131 | ############ | |
132 | ||
133 | ∇₂ᵀ!(Δx, y) # primal step: | |
134 | @. x̄ = x # | save old x for over-relax | |
135 | @. x = (x-τ*(Δx-b))/(1+τ) # | prox | |
136 | horn_schunck_reg_prox!(u, b_next_filt, b_filt, θ, λ, T, τ) | |
137 | @. x̄ = 2x - x̄ # over-relax | |
138 | ∇₂!(y, x̄) # dual step: y | |
139 | @. y = (y + σ*Δy)/(1 + σ*ρ/α) # | | |
140 | proj_norm₂₁ball!(y, α) # | prox | |
141 | ||
142 | if DisplacementT == DisplacementConstant | |
143 | ucumul .+= u | |
144 | end | |
145 | ||
146 | ######################################################## | |
147 | # Give function value and cumulative movement if needed | |
148 | ######################################################## | |
149 | v = verbose() do | |
150 | ∇₂!(Δy, x) | |
151 | tmp = zeros(size(b_filt)) | |
152 | pointwise_gradiprod_2d!(tmp, Δy, u, b_filt) | |
153 | value = (norm₂²(b-x)/2 + θ*norm₂²((b_next_filt-b_filt)./T+tmp) | |
154 | + λ*norm₂²(u)/2 + α*γnorm₂₁(Δy, ρ)) | |
155 | ||
156 | value, x, ucumul, nothing | |
157 | end | |
158 | ||
159 | return v | |
160 | end | |
161 | ||
162 | return x, y, v | |
163 | end | |
164 | ||
165 | end # Module | |
166 | ||
167 |