Test Registration Functions
In this example page I will show how the registration algorithms (lap, pflap, sparse_lap and sparse_pflap) perform. I will will show their speed and the quality of their outputs.
First I some testing images:
# using the package and TimerOutputs
using LAP_julia, TimerOutputs
# default arguments:
img, imgw, flow = gen_init();See the Basic Interaction section or the Public Documentation for more advanced input generation
These are the differences between the target (img) and source (imgw) images.
imgoverlay(img, imgw, figtitle="Target vs Source")
test_registration_alg
This function times the registration algorithm on the inside and then test the output flow - flow_est and aligned source - source_reg.
# choose a method
method = sparse_pflap
# start a timer
timer = TimerOutput("ALG: " * string(method))
# set the keyword arguments of the method
method_kwargs = Dict(:display => false, :timer => timer, :match_source_histogram => false)
# set the arguments of the method if any
method_args = []
# start the test
flow_est, source_reg, timer, results = test_registration_alg(method,
img,
imgw,
flow,
timer=timer,
method_args=method_args,
only_flow_compare=false, # this adds source reg tests.
method_kwargs=method_kwargs); ──────────────────────────────────────────────────────────────────────────────
Time Allocations
────────────────────── ───────────────────────
Tot / % measured: 353ms / 96.3% 167MiB / 96.3%
Section ncalls time %tot avg alloc %tot avg
──────────────────────────────────────────────────────────────────────────────
ALG: sparse_pflap 1 340ms 100% 340ms 161MiB 100% 161MiB
single filter py... 7 321ms 94.6% 45.9ms 154MiB 96.2% 22.1MiB
single lap at ... 7 287ms 84.6% 41.1ms 121MiB 75.6% 17.4MiB
filtering 7 238ms 70.2% 34.1ms 68.2MiB 42.5% 9.75MiB
prepare A and b 7 26.3ms 7.73% 3.75ms 41.7MiB 25.9% 5.95MiB
window sum... 21 15.6ms 4.59% 743μs 25.0MiB 15.6% 1.19MiB
window sum... 14 10.6ms 3.13% 759μs 16.6MiB 10.4% 1.19MiB
multi mat div 7 120μs 0.04% 17.2μs 298KiB 0.18% 42.5KiB
calculate flow 7 26.9μs 0.01% 3.84μs 49.9KiB 0.03% 7.13KiB
image interpol... 7 19.3ms 5.69% 2.76ms 21.0MiB 13.1% 3.00MiB
flow interpola... 7 13.8ms 4.05% 1.97ms 11.7MiB 7.28% 1.67MiB
filter inds 7 272μs 0.08% 38.9μs 245KiB 0.15% 35.0KiB
setup 1 18.1ms 5.34% 18.1ms 6.13MiB 3.82% 6.13MiB
find edge points 1 17.2ms 5.06% 17.2ms 4.13MiB 2.57% 4.13MiB
──────────────────────────────────────────────────────────────────────────────
mse | 0.001
rmse | 0.023
time | 0.34
ncc | 0.993
flow_mae | 0.144
angle-rmse | 30.816
angle-mae | 4.322
mae | 0.012
flow_rmse | 0.032Check the generated flow and the original flow:
Ground Truth:
showflow(flow.*(-1), figtitle="Ground Truth")
Estimated Flow:
showflow(flow_est, figtitle="Estimated Flow")
Compare it with the pflap function
method = pflap
# same inputs, new timer
timer = TimerOutput("ALG: " * string(method))
flow_est, source_reg, timer, results = test_registration_alg(method,
img,
imgw,
flow,
timer=timer,
method_args=method_args,
only_flow_compare=false, # this adds source reg tests.
method_kwargs=method_kwargs); ─────────────────────────────────────────────────────────────────────
Time Allocations
────────────────────── ───────────────────────
Tot / % measured: 1.26s / 99.1% 692MiB / 99.1%
Section ncalls time %tot avg alloc %tot avg
─────────────────────────────────────────────────────────────────────
ALG: pflap 1 1.25s 100% 1.25s 686MiB 100% 686MiB
─────────────────────────────────────────────────────────────────────
mse | 0.0
rmse | 0.022
time | 1.246
ncc | 0.994
flow_mae | 0.214
angle-rmse | 20.649
angle-mae | 3.722
mae | 0.012
flow_rmse | 0.083Check the generated flow and the original flow:
Ground Truth:
showflow(flow.*(-1), figtitle="Ground Truth")Estimated Flow:
showflow(flow_est, figtitle="Estimated Flow")
The time_reg_alg is used to time the registration algorithm, when the original flow isn't known.
using the display keyword
The registration algorithms have a display keyword argument, that can be set to true to output figures showing the flow_est at each iteration and print some extra debug info.
method = sparse_pflap
timer = TimerOutput("ALG: " * string(method))
method_kwargs = Dict(:display => true, :timer => timer, :match_source_histogram => false)
method_args = []
flow_est, source_reg, timer, results, (figs,) = test_registration_alg(method,
img,
imgw,
flow,
timer=timer,
method_args=method_args,
only_flow_compare=false, # this adds source reg tests.
method_kwargs=method_kwargs);ind count: 297
###################
ITERATION: 1
filter_half_size: 64
inner loop iter: 1
current inds: 69
non NaN new estim flow vector count: 69
###################
ITERATION: 2
filter_half_size: 32
inner loop iter: 1
current inds: 157
non NaN new estim flow vector count: 157
###################
ITERATION: 3
filter_half_size: 16
inner loop iter: 1
current inds: 210
non NaN new estim flow vector count: 210
###################
ITERATION: 4
filter_half_size: 8
inner loop iter: 1
current inds: 243
non NaN new estim flow vector count: 243
###################
ITERATION: 5
filter_half_size: 4
inner loop iter: 1
current inds: 256
non NaN new estim flow vector count: 256
###################
ITERATION: 6
filter_half_size: 2
inner loop iter: 1
current inds: 272
non NaN new estim flow vector count: 268
###################
ITERATION: 7
filter_half_size: 1
inner loop iter: 1
current inds: 282
non NaN new estim flow vector count: 245
──────────────────────────────────────────────────────────────────────────────
Time Allocations
────────────────────── ───────────────────────
Tot / % measured: 1.26s / 99.0% 406MiB / 98.5%
Section ncalls time %tot avg alloc %tot avg
──────────────────────────────────────────────────────────────────────────────
ALG: sparse_pflap 1 1.25s 100% 1.25s 400MiB 100% 400MiB
single filter py... 7 1.23s 98.7% 176ms 394MiB 98.5% 56.2MiB
single lap at ... 7 284ms 22.8% 40.6ms 121MiB 30.4% 17.4MiB
filtering 7 257ms 20.6% 36.7ms 68.2MiB 17.1% 9.75MiB
prepare A and b 7 26.8ms 2.15% 3.83ms 41.7MiB 10.4% 5.95MiB
window sum... 21 15.8ms 1.27% 752μs 25.0MiB 6.26% 1.19MiB
window sum... 14 11.0ms 0.88% 786μs 16.6MiB 4.16% 1.19MiB
multi mat div 7 132μs 0.01% 18.8μs 298KiB 0.07% 42.5KiB
calculate flow 7 26.4μs 0.00% 3.77μs 49.9KiB 0.01% 7.13KiB
image interpol... 7 30.4ms 2.44% 4.34ms 21.0MiB 5.25% 3.00MiB
flow interpola... 7 16.1ms 1.29% 2.30ms 11.7MiB 2.93% 1.67MiB
filter inds 7 266μs 0.02% 38.1μs 245KiB 0.06% 35.0KiB
setup 1 16.0ms 1.28% 16.0ms 6.13MiB 1.53% 6.13MiB
find edge points 1 15.2ms 1.22% 15.2ms 4.13MiB 1.03% 4.13MiB
──────────────────────────────────────────────────────────────────────────────
mse | 0.001
rmse | 0.023
time | 1.246
ncc | 0.993
flow_mae | 0.144
angle-rmse | 30.816
angle-mae | 4.322
mae | 0.012
flow_rmse | 0.032This page was generated using Literate.jl.