(0016799)
Luis Ibanez (manager)
2009-07-01 15:03
edited on: 2009-07-01 15:08

The first suspect was the actual list of points (indexes) selected by the random iterator.

However, I have verified that the set of indexes is the same, regardless of how many threads are set in the Metric.

In the same exercise, by running the existing tests:

itkOptImageToImageMetricsTest01
itkOptImageToImageMetricsTest02
itkOptImageToImageMetricsTest03

I have verified that the values of the derivative change
depending on the number of threads used in the computation.]
(which is an indication of a bug).

(0016800)
Hans Johnson (developer)
2009-07-01 16:54

Luis,

It was also noted during NAMIC that doing a rigid registration where the fixed and moving images are exactly the same image does not result in an identity matrix. It is hypothesized that this is also a problem with the derivatives not being computed correctly.

Perhaps doing a test with exactly the same image as fixed and moving will help you identify the problem more easily.

Hans

(0016801)
Luis Ibanez (manager)
2009-07-01 16:57

Hans,

Thanks for pointing this out.

This new test is actually using two identical images,
but it is not performing a full registration, only
calls to GetValue() and GetValueAndDerivative().

This should still give us some hints on where the
source of the problem may be.

We probably should add a couple of tests with the
situation that you describe. To verify that a
registration against an identical image results
in Identity transforms.

Thanks

Luis

(0016802)
Luis Ibanez (manager)
2009-07-01 17:54

The same problem happens for the multi-threaded MeanSquares metric.
That is, the values returned by GetValue() and GetValueAndDerivative() change
depending on the number of threads used.

In MeanSquares, we have verified that the metric value returned by GetValue() is the same that the one returned by GetValueAndDerivative(), as well as verified that the value returned by GetDerivative() is the same as the one returned by GetValueAndDerivative().

(0016803)
Luis Ibanez (manager)
2009-07-01 17:55

a test has been added to the Nightly suite.

(0016804)
Luis Ibanez (manager)
2009-07-01 22:45

After adding a proper Reinitialization of the Seed (for the random pixel iterator), the results are the same in multiple threads when using less than 50,000 samples.

At 50,000 samples, on the other hand, the values of metric and derivative start
to diverge in the low digits. It start looking like a progressive accumulation of errors...

(0022980)
Luis Ibanez (manager)
2010-11-07 01:58

This should be addressed as part of the image registration refactoring in ITKv4

Issue History
Date Modified	Username	Field	Change
2009-06-30 18:15	Luis Ibanez	New Issue
2009-06-30 18:15	Luis Ibanez	Status	new => assigned
2009-06-30 18:15	Luis Ibanez	Assigned To	=> Luis Ibanez
2009-07-01 15:03	Luis Ibanez	Note Added: 0016799
2009-07-01 15:08	Luis Ibanez	Note Edited: 0016799
2009-07-01 16:54	Hans Johnson	Note Added: 0016800
2009-07-01 16:57	Luis Ibanez	Note Added: 0016801
2009-07-01 17:54	Luis Ibanez	Note Added: 0016802
2009-07-01 17:55	Luis Ibanez	Note Added: 0016803
2009-07-01 17:55	Luis Ibanez	Status	assigned => confirmed
2009-07-01 17:55	Luis Ibanez	Priority	normal => high
2009-07-01 22:45	Luis Ibanez	Note Added: 0016804
2010-11-07 01:58	Luis Ibanez	Status	confirmed => assigned
2010-11-07 01:58	Luis Ibanez	Assigned To	Luis Ibanez => brian avants
2010-11-07 01:58	Luis Ibanez	Note Added: 0022980

Notes
(0016799) Luis Ibanez (manager) 2009-07-01 15:03 edited on: 2009-07-01 15:08	The first suspect was the actual list of points (indexes) selected by the random iterator. However, I have verified that the set of indexes is the same, regardless of how many threads are set in the Metric. In the same exercise, by running the existing tests: itkOptImageToImageMetricsTest01 itkOptImageToImageMetricsTest02 itkOptImageToImageMetricsTest03 I have verified that the values of the derivative change depending on the number of threads used in the computation.] (which is an indication of a bug).

(0016800) Hans Johnson (developer) 2009-07-01 16:54	Luis, It was also noted during NAMIC that doing a rigid registration where the fixed and moving images are exactly the same image does not result in an identity matrix. It is hypothesized that this is also a problem with the derivatives not being computed correctly. Perhaps doing a test with exactly the same image as fixed and moving will help you identify the problem more easily. Hans

(0016801) Luis Ibanez (manager) 2009-07-01 16:57	Hans, Thanks for pointing this out. This new test is actually using two identical images, but it is not performing a full registration, only calls to GetValue() and GetValueAndDerivative(). This should still give us some hints on where the source of the problem may be. We probably should add a couple of tests with the situation that you describe. To verify that a registration against an identical image results in Identity transforms. Thanks Luis

(0016802) Luis Ibanez (manager) 2009-07-01 17:54	The same problem happens for the multi-threaded MeanSquares metric. That is, the values returned by GetValue() and GetValueAndDerivative() change depending on the number of threads used. In MeanSquares, we have verified that the metric value returned by GetValue() is the same that the one returned by GetValueAndDerivative(), as well as verified that the value returned by GetDerivative() is the same as the one returned by GetValueAndDerivative().

(0016803) Luis Ibanez (manager) 2009-07-01 17:55	a test has been added to the Nightly suite.

(0016804) Luis Ibanez (manager) 2009-07-01 22:45	After adding a proper Reinitialization of the Seed (for the random pixel iterator), the results are the same in multiple threads when using less than 50,000 samples. At 50,000 samples, on the other hand, the values of metric and derivative start to diverge in the low digits. It start looking like a progressive accumulation of errors...

(0022980) Luis Ibanez (manager) 2010-11-07 01:58	This should be addressed as part of the image registration refactoring in ITKv4

Relationships