Tue, 07 Apr 2020 14:19:48 -0500
Initialise independent repo
0 | 1 | ###################################################################### |
2 | # Predictive online PDPS for optical flow with unknown velocity field | |
3 | ###################################################################### | |
4 | ||
5 | __precompile__() | |
6 | ||
7 | module AlgorithmBothMulti | |
8 | ||
9 | using Printf | |
10 | ||
11 | using AlgTools.Util | |
12 | import AlgTools.Iterate | |
13 | using ImageTools.Gradient | |
14 | using ImageTools.ImFilter | |
15 | ||
16 | using ..OpticalFlow: Image, | |
17 | ImageSize, | |
18 | DisplacementConstant, | |
19 | pdflow!, | |
20 | horn_schunck_reg_prox!, | |
21 | pointwise_gradiprod_2d!, | |
22 | horn_schunck_reg_prox_op!, | |
23 | mldivide_step_plus_sym2x2!, | |
24 | ConstantDisplacementHornSchunckData, | |
25 | filter_hs | |
26 | ||
27 | using ..Algorithm: step_lengths | |
28 | ||
29 | ######################### | |
30 | # Iterate initialisation | |
31 | ######################### | |
32 | ||
33 | function init_displ(xinit::Image, ::Type{DisplacementConstant}, n::Integer) | |
34 | return xinit, zeros(n, 2) | |
35 | end | |
36 | ||
37 | # function init_displ(xinit::Image, ::Type{DisplacementFull}, n::Integer) | |
38 | # return xinit, zeros(n, 2, size(xinit)...) | |
39 | # end | |
40 | ||
41 | function init_rest(x::Image, u::DisplacementT) where DisplacementT | |
42 | imdim=size(x) | |
43 | ||
44 | y = zeros(2, imdim...) | |
45 | Δx = copy(x) | |
46 | Δy = copy(y) | |
47 | x̄ = copy(x) | |
48 | ||
49 | return x, y, Δx, Δy, x̄, u | |
50 | end | |
51 | ||
52 | function init_iterates( :: Type{DisplacementT}, | |
53 | xinit::Image, | |
54 | n::Integer) where DisplacementT | |
55 | return init_rest(init_displ(copy(xinit), DisplacementT, n)...) | |
56 | end | |
57 | ||
58 | function init_iterates( :: Type{DisplacementT}, | |
59 | dim::ImageSize, | |
60 | n::Integer) where DisplacementT | |
61 | return init_rest(init_displ(zeros(dim...), DisplacementT, n)...) | |
62 | end | |
63 | ||
64 | ############ | |
65 | # Algorithm | |
66 | ############ | |
67 | ||
68 | function solve( :: Type{DisplacementT}; | |
69 | dim :: ImageSize, | |
70 | iterate = AlgTools.simple_iterate, | |
71 | params::NamedTuple) where DisplacementT | |
72 | ||
73 | ###################### | |
74 | # Initialise iterates | |
75 | ###################### | |
76 | ||
77 | n = params.displacement_count | |
78 | k = 0 # number of displacements we have already | |
79 | ||
80 | x, y, Δx, Δy, x̄, u = init_iterates(DisplacementT, dim, n) | |
81 | init_data = (params.init == :data) | |
82 | hs = [ConstantDisplacementHornSchunckData() for i=1:n] | |
83 | #hs = Array{ConstantDisplacementHornSchunckData}(undef, n) | |
84 | A = Array{Float64,3}(undef, n, 2, 2) | |
85 | d = Array{Float64,2}(undef, n, 2) | |
86 | ||
87 | # … for tracking cumulative movement | |
88 | ucumulbase = [0.0, 0.0] | |
89 | ||
90 | ############################################# | |
91 | # Extract parameters and set up step lengths | |
92 | ############################################# | |
93 | ||
94 | α, ρ, λ, θ = params.α, params.ρ, params.λ, params.θ | |
95 | R_K² = ∇₂_norm₂₂_est² | |
96 | γ = 1 | |
97 | τ, σ, σ̃, ρ̃ = step_lengths(params, γ, R_K²) | |
98 | ||
99 | kernel = params.kernel | |
100 | T = params.timestep | |
101 | ||
102 | #################### | |
103 | # Run the algorithm | |
104 | #################### | |
105 | ||
106 | b_next_filt = nothing | |
107 | diffu = similar(u[1, :]) | |
108 | ||
109 | v = iterate(params) do verbose :: Function, | |
110 | b :: Image, | |
111 | 🚫unused_v_known :: DisplacementT, | |
112 | b_next :: Image | |
113 | ||
114 | #################################### | |
115 | # Smooth data for Horn–Schunck term | |
116 | #################################### | |
117 | ||
118 | b_filt, b_next_filt = filter_hs(b, b_next, b_next_filt, kernel) | |
119 | ||
120 | ################################################ | |
121 | # Prediction step | |
122 | ################################################ | |
123 | ||
124 | # Predict x and y | |
125 | if k==0 | |
126 | if init_data | |
127 | x .= b | |
128 | init_data = false | |
129 | end | |
130 | else | |
131 | # Displacement from previous to this image is estimated as | |
132 | # the difference of their displacements from beginning of window. | |
133 | if k>1 | |
134 | @. @views diffu = u[k, :] - u[k-1, :] | |
135 | else | |
136 | @. @views diffu = u[k, :] | |
137 | end | |
138 | ||
139 | pdflow!(x, Δx, y, Δy, diffu, params.dual_flow) | |
140 | end | |
141 | ||
142 | # Shift stored prox matrices | |
143 | if k==n | |
144 | tmp = copy(u[1, :]) | |
145 | ucumulbase .+= tmp | |
146 | for j=1:(n-1) | |
147 | @. @views u[j, :] = u[j+1, :] - tmp | |
148 | hs[j] = hs[j+1] | |
149 | end | |
150 | # Create new struct as original contains references to objects that | |
151 | # have been moved to index n-1. | |
152 | hs[n]=ConstantDisplacementHornSchunckData() | |
153 | else | |
154 | k += 1 | |
155 | end | |
156 | ||
157 | # Predict u: zero displacement from current to next image, i.e., | |
158 | # same displacement to beginning of window. | |
159 | if k==1 | |
160 | @. @views u[k, :] = 0.0 | |
161 | else | |
162 | @. @views u[k, :] = u[k-1, :] | |
163 | end | |
164 | ||
165 | # Predictor proximals tep | |
166 | if params.prox_predict | |
167 | ∇₂!(Δy, x) | |
168 | @. y = (y + σ̃*Δy)/(1 + σ̃*(ρ̃+ρ/α)) | |
169 | proj_norm₂₁ball!(y, α) | |
170 | end | |
171 | ||
172 | ################################################################################# | |
173 | # PDPS step | |
174 | # | |
175 | # For the displacements, with τ̃=τ/k, we need to solve for 2≤j<k, | |
176 | # | |
177 | # (1) (I/τ̃+M₀^j+M₀^{j+1})u^j = M₀^ju^{j-1} + M₀^{j+1}u^{j+1} | |
178 | # + ũ^j/τ̃ - z^j + z^{j+1}, | |
179 | # | |
180 | # as well as | |
181 | # | |
182 | # (2) (I/τ̃+M₀^k)u^k = M₀^k u^{k-1} + ũ^k/τ̃ - z^k | |
183 | # | |
184 | # and | |
185 | # | |
186 | # (3) (I/τ̃+M₀^1+M₀^2)u^1 = 0 + M₀^{2}u^{2} + ũ^1/τ̃ - z^1 + z^{2} | |
187 | # | |
188 | # We first construct from (2) that | |
189 | # | |
190 | # u^k = A^k u^{k-1} + d^k | |
191 | # | |
192 | # for | |
193 | # | |
194 | # A^k := (I/τ̃+M₀^k)^{-1} M₀^k | |
195 | # d_k := (I/τ̃+M₀^k)^{-1} (ũ^k/τ̃ - z^k). | |
196 | # | |
197 | # Inserting this into (1) we need | |
198 | # | |
199 | # (4) (I/τ̃+M₀^j+M₀^{j+1}(I-A^{j+1}))u^j = M₀^ju^{j-1} + M₀^{j+1}d^{j+1} | |
200 | # + ũ^j/τ̃ - z^j + z^{j+1}. | |
201 | # | |
202 | # This is well-defined because A^{j+1} < I. It also has the same form as (1), so | |
203 | # we continue with | |
204 | # | |
205 | # (5) u^j = A^j u^{j-1} + d^j | |
206 | # | |
207 | # for | |
208 | # | |
209 | # A^j := (I/τ̃+M₀^j+M₀^{j+1}(I-A^{j+1}))^{-1} M₀^j | |
210 | # d^j := (I/τ̃+M₀^j+M₀^{j+1}(I-A^{j+1}))^{-1} | |
211 | # (M₀^{j+1}d^{j+1} + ũ^j/τ̃ - z^j + z^{j+1}) | |
212 | # | |
213 | # Finally from (3) with these we need | |
214 | # | |
215 | # (I/τ̃+M₀^1+M₀^2(I-A^2))u^1 = M₀^2d^2 + ũ^1/τ̃ - z^1 + z^2, | |
216 | # | |
217 | # which is of the same form as (4) with u^0=0, so by (5) u^1=d^1. | |
218 | # | |
219 | ################################################################################# | |
220 | ||
221 | ∇₂ᵀ!(Δx, y) # primal step: | |
222 | @. x̄ = x # | save old x for over-relax | |
223 | @. x = (x-τ*(Δx-b))/(1+τ) # | prox | |
224 | # | | for displacement | |
225 | # Calculate matrices for latest data; rest is stored. | |
226 | @views begin | |
227 | horn_schunck_reg_prox_op!(hs[k], b_next_filt, b_filt, θ, λ, T) | |
228 | ||
229 | τ̃=τ/k | |
230 | ||
231 | B = hs[k].M₀ | |
232 | c = u[k, :]./τ̃-hs[k].z | |
233 | mldivide_step_plus_sym2x2!(A[k, :, :], B, B, τ̃) | |
234 | mldivide_step_plus_sym2x2!(d[k, :], B, c, τ̃) | |
235 | ||
236 | for j=(k-1):-1:1 | |
237 | B = hs[j].M₀+hs[j+1].M₀*([1 0; 0 1]-A[j+1, :, :]) | |
238 | c = hs[j+1].M₀*d[j+1, :]+u[j, :]./τ̃-hs[j].z+hs[j+1].z | |
239 | mldivide_step_plus_sym2x2!(A[j, :, :], B, hs[j].M₀, τ̃) | |
240 | mldivide_step_plus_sym2x2!(d[j, :], B, c, τ̃) | |
241 | end | |
242 | ||
243 | u[1, :] .= d[1, :] | |
244 | for j=2:k | |
245 | u[j, :] .= A[j, :, :]*u[j-1, :] + d[j, :] | |
246 | end | |
247 | end | |
248 | ||
249 | @. x̄ = 2x - x̄ # over-relax: x̄ = 2x-x_old | |
250 | ∇₂!(Δy, x̄) # dual step: y | |
251 | @. y = (y + σ*Δy)/(1 + σ*ρ/α) # | | |
252 | proj_norm₂₁ball!(y, α) # | prox | |
253 | ||
254 | ######################################################## | |
255 | # Give function value and cumulative movement if needed | |
256 | ######################################################## | |
257 | v = verbose() do | |
258 | ∇₂!(Δy, x) | |
259 | hs_plus_reg=0 | |
260 | for j=1:k | |
261 | v=(j==1 ? u[j, :] : u[j, :]-u[j-1, :]) | |
262 | hs_plus_reg += hs[j].cv/2 + dot(hs[j].Mv*v, v)/2+dot(hs[j].av, v) | |
263 | end | |
264 | value = (norm₂²(b-x)/2 + hs_plus_reg/k + α*γnorm₂₁(Δy, ρ)) | |
265 | ||
266 | value, x, u[k, :]+ucumulbase, u[1:k,:].+ucumulbase' | |
267 | end | |
268 | ||
269 | return v | |
270 | end | |
271 | ||
272 | return x, y, v | |
273 | end | |
274 | ||
275 | end # Module | |
276 | ||
277 |