Seperable Convolution with Asymmetric Kernels + Fast Gauss #46

maarzt · 2017-10-09T14:10:18Z

Separable asymmetric kernels are used in image processing.
One example: The derivative of a gauss kernel is a separable, asymmetric kernel, and is used to calculate exact derivatives of an gaussian blurred image.

This PR deprecates the current separable symmetric convolution algorithm. And adds generalized algorithm for separable convolution. The new algorithm has been benchmarked against the current implementation an proved to be just as fast.

Gauss3 is changed to use the new implementation.

maarzt · 2017-10-09T14:13:05Z

Do we want to move the classes for separable (symmetric) convolution to net.imglib2.algorithm.convolution? (done)

maarzt · 2018-05-01T01:02:55Z

I think this PR is ready to be merged and I need it to implement features for pixel classification.

This makes ClassCopyProvider available.

tpietzsch · 2018-08-16T13:34:25Z

src/main/java/net/imglib2/algorithm/convolution/fast_gauss/FastGauss.java

+ * <p>
+ * The result is a good approximation to "real" gauss. But the
+ * same holds for {@link net.imglib2.algorithm.gauss3.Gauss3}.
+ * Which one is better is not known so far.


This comment is not very helpful IMO. Would be better to state what the difference in approximation is.
I.e. Gauss3 does a (non-approximated) convolution with a Gaussian kernel truncated at 3*sigma+1.
What is the approximation here?

Also, you should add a reference to the paper that is implemented here.

tpietzsch · 2018-08-16T13:35:27Z

src/main/java/net/imglib2/algorithm/gauss3/Gauss3.java

 		final double[][] halfkernels = halfkernels( sigma );
 		final ExecutorService service = Executors.newFixedThreadPool( numThreads );
-		SeparableSymmetricConvolution.convolve( halfkernels, source, target, service );
+		SeparableKernelConvolution.convolve( Kernel1D.symmetric( halfkernels ), source, target );


This does not use the provided ExecutorService

tpietzsch · 2018-08-16T13:35:36Z

src/main/java/net/imglib2/algorithm/gauss3/Gauss3.java

 	{
 		final double[][] halfkernels = halfkernels( sigma );
-		SeparableSymmetricConvolution.convolve( halfkernels, source, target, service );
+		SeparableKernelConvolution.convolve( Kernel1D.symmetric( halfkernels ), source, target );


This does not use the provided ExecutorService

tpietzsch · 2018-08-16T13:41:09Z

src/main/java/net/imglib2/algorithm/convolution/kernel/SeparableKernelConvolution.java

+	/**
+	 * Return an object, that performs the separable convolution with the give kernel.
+	 * It additional allows to set the {@link java.util.concurrent.ExecutorService} to use
+	 * or to query for the required input image size, ...


Typos: "give" --> "given"
"additional" --> "additionally"

It's faster than the older one and allows asymmetric kernels.

It's functionality is replaced by SeparableKernelConvolution.

It is an approximate algorithm to calculate Gaussian blur. But way faster for sigma greater than 3.

tpietzsch · 2018-08-27T13:22:58Z

@maarzt Could you please fix, what I have commented on so far?

maarzt · 2018-08-29T10:05:35Z

@tpietzsch I made all the changes you required. And finally I cited the paper describing the fast gauss algorithm that is implemented.

tpietzsch · 2018-08-29T10:39:38Z

@maarzt Did you push your changes? Gauss3 still doesn't use user-provided ExecutorService

…torService

* Use fork( 1 ) to avoid potential dependency of results on execution order. * Longer warmup for GaussBenchmark.

This makes it clear that benchmarkSeparableKernelConvolution() is using the same number of threads as benchmarkSeparableSymmetricConvolution().

…lvers

Also brings a slight speed up

The tests are no longer needed, because they test the same functionality as ConvolverTest.

…e except DoubleType This is required for higher precision when blurring for example ByteType images.

maarzt · 2018-10-17T14:45:40Z

@tpietzsch I did the changes you asked for:

Intermediate results for convolution of all RealTypes except DoubleType is now FloatType.
I use assertImageEquals from new imglib2-5.6.0

tpietzsch · 2018-10-18T12:11:01Z

Great, thanks!

maarzt force-pushed the seperable-convolution branch 3 times, most recently from f190097 to a2b3399 Compare November 9, 2017 09:44

Add JMH dependency whit scope test

134099e

maarzt requested a review from tpietzsch May 1, 2018 01:03

maarzt force-pushed the seperable-convolution branch from a2b3399 to c2718c2 Compare May 1, 2018 01:12

maarzt changed the title ~~Seperable Convolution with Asymmetric Kernels~~ Seperable Convolution with Asymmetric Kernels (depends on https://github.com/imglib/imglib2/pull/206) May 2, 2018

maarzt changed the title ~~Seperable Convolution with Asymmetric Kernels (depends on https://github.com/imglib/imglib2/pull/206)~~ Seperable Convolution with Asymmetric Kernels (imglib2 PR #206) May 2, 2018

maarzt mentioned this pull request May 2, 2018

Make ClassCopyProvider public, I would like to use it in imglib2-algorithms imglib/imglib2#206

Merged

maarzt added 2 commits July 30, 2018 15:22

Add test for SeparableSymmetricConvolution

718f526

Update parent to pom-imglib2 11.0.1

9a4725d

This makes ClassCopyProvider available.

maarzt force-pushed the seperable-convolution branch from 43b47b5 to 574e43e Compare July 30, 2018 13:36

maarzt changed the title ~~Seperable Convolution with Asymmetric Kernels (imglib2 PR #206)~~ Seperable Convolution with Asymmetric Kernels + Fast Gauss Jul 30, 2018

tpietzsch reviewed Aug 16, 2018

View reviewed changes

maarzt added 5 commits August 20, 2018 15:46

Add package for separable convolution

e3eb043

It's faster than the older one and allows asymmetric kernels.

Use new separable convolution to calculate Gauss3

1aa76a5

Deprecate SeparableSymmetricConvolution

14937f4

It's functionality is replaced by SeparableKernelConvolution.

Add fast gauss implementation

8b3a43d

It is an approximate algorithm to calculate Gaussian blur. But way faster for sigma greater than 3.

Add Benchmark to compare performance of convolution implementations

64429bb

maarzt force-pushed the seperable-convolution branch from 574e43e to 64429bb Compare August 20, 2018 13:49

FastGauss: improve javadoc

6d3d0dd

ConvolverBenchmark: add benchmarks for non double convolvers

44d2c88

maarzt and others added 16 commits August 31, 2018 15:47

AbstractMultiThreadedConvolution: shutdown automatically create Execu…

c619131

…torService

Fix typos

15a694e

Fix benchmark parameters.

ec75625

* Use fork( 1 ) to avoid potential dependency of results on execution order. * Longer warmup for GaussBenchmark.

Use explicit ExecutorService in GaussBenchmark.

b49cbf5

This makes it clear that benchmarkSeparableKernelConvolution() is using the same number of threads as benchmarkSeparableSymmetricConvolution().

Don't reimplement net.imglib2.util.Util.getArrayOrCellImgFactory

5b07795

Reuse Intervals.numElements

54a18da

Basic test for Convolver variants (NativeType, NumericType, etc)

da903dd

For readability, use the same names and access modifiers in all convo…

f46e9bd

…lvers

Set buffer tail to zero just to be sure

85d9b2b

Make ConvolverNativeType.process logic the same as for other convolvers

b740bad

Also brings a slight speed up

Remove unused method

ac0dbee

Formatting

b14b020

ConvolverTest: reduce code duplication

63b4ea9

KernelConvolverFactoryTest: remove redundant tests

e48dd73

The tests are no longer needed, because they test the same functionality as ConvolverTest.

KernelConvolverFactory: Use FloatType for all convolutions on RealTyp…

65abb18

…e except DoubleType This is required for higher precision when blurring for example ByteType images.

Use assertImageEquals when possible

d17d33e

tpietzsch merged commit de30b1a into imglib:master Oct 18, 2018

maarzt deleted the seperable-convolution branch August 27, 2019 08:57

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Seperable Convolution with Asymmetric Kernels + Fast Gauss #46

Seperable Convolution with Asymmetric Kernels + Fast Gauss #46

Uh oh!

maarzt commented Oct 9, 2017 •

edited

Loading

Uh oh!

maarzt commented Oct 9, 2017 •

edited

Loading

Uh oh!

maarzt commented May 1, 2018

Uh oh!

tpietzsch Aug 16, 2018

Uh oh!

tpietzsch Aug 16, 2018

Uh oh!

tpietzsch Aug 16, 2018

Uh oh!

tpietzsch Aug 16, 2018

Uh oh!

tpietzsch commented Aug 27, 2018

Uh oh!

maarzt commented Aug 29, 2018

Uh oh!

tpietzsch commented Aug 29, 2018

Uh oh!

maarzt commented Oct 17, 2018

Uh oh!

tpietzsch commented Oct 18, 2018

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Seperable Convolution with Asymmetric Kernels + Fast Gauss #46

Seperable Convolution with Asymmetric Kernels + Fast Gauss #46

Uh oh!

Conversation

maarzt commented Oct 9, 2017 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

maarzt commented Oct 9, 2017 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

maarzt commented May 1, 2018

Uh oh!

tpietzsch Aug 16, 2018

Choose a reason for hiding this comment

Uh oh!

tpietzsch Aug 16, 2018

Choose a reason for hiding this comment

Uh oh!

tpietzsch Aug 16, 2018

Choose a reason for hiding this comment

Uh oh!

tpietzsch Aug 16, 2018

Choose a reason for hiding this comment

Uh oh!

tpietzsch commented Aug 27, 2018

Uh oh!

maarzt commented Aug 29, 2018

Uh oh!

tpietzsch commented Aug 29, 2018

Uh oh!

maarzt commented Oct 17, 2018

Uh oh!

tpietzsch commented Oct 18, 2018

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

maarzt commented Oct 9, 2017 •

edited

Loading

maarzt commented Oct 9, 2017 •

edited

Loading