COMP: Don't let ImageToFeaturesMap output inherit the input LargestPossibleRegion

ImageToFeaturesMap is an itk::ProcessObject whose outputs are feature maps whose
size and spacing are data-dependent (decided by the model) and set in
GenerateData() by grafting the TensorToImageFilter result. The default
ProcessObject::GenerateOutputInformation() copies the primary input's geometry --
including its LargestPossibleRegion -- onto every output. When a downstream filter
re-propagates UpdateOutputInformation() up this filter (e.g. the per-channel
VectorIndexSelectionCast in InterpolateVectorImageFunction::SetInputImage), that
default resets each output's LargestPossibleRegion back to the larger input image
size while the buffer still holds the smaller feature map, so the consumer then
iterates an input-sized region over the smaller buffer and throws
'Region ... is outside of buffered region'.

Override GenerateOutputInformation() with an empty body so the grafted regions are
never clobbered. Surfaced by the Elastix IMPACT metric on static-ITK builds; the
Advanced (Elastix) and v4 (Python) metrics both consume this filter.

Author: vboussot

include/itkImageToFeaturesMap.h

@@ -167,6 +167,23 @@ class ImageToFeaturesMap : public ProcessObject
   void
   VerifyPreconditions() const override;
 
+  /** Deliberately suppress the default output-information generation.
+   *
+   * ProcessObject::GenerateOutputInformation() copies the primary input's geometry
+   * (origin/spacing/direction AND LargestPossibleRegion) onto every output. That is
+   * wrong for this filter: the outputs are feature maps whose size and spacing are
+   * data-dependent (decided by the model) and are set authoritatively in
+   * GenerateData() by grafting the TensorToImageFilter result. Worse, when a
+   * downstream filter re-propagates UpdateOutputInformation() up this filter (e.g. the
+   * per-channel VectorIndexSelectionCast in InterpolateVectorImageFunction::SetInputImage),
+   * the default would reset each output's LargestPossibleRegion back to the (larger)
+   * input image size while the buffer still holds the smaller feature map -- the
+   * consumer then iterates an input-sized region over the smaller buffer and throws
+   * "Region ... is outside of buffered region". Overriding with an empty body keeps the
+   * grafted regions intact. */
+  void
+  GenerateOutputInformation() override;
+
   void
   GenerateData() override;
 

include/itkImageToFeaturesMap.hxx

@@ -80,6 +80,19 @@ ImageToFeaturesMap<TInputImage, TInterpolator>::VerifyPreconditions() const
   Superclass::VerifyPreconditions();
 }
 
+template <typename TInputImage, typename TInterpolator>
+void
+ImageToFeaturesMap<TInputImage, TInterpolator>::GenerateOutputInformation()
+{
+  // Intentionally empty: do NOT chain to Superclass (ProcessObject), whose default
+  // copies the primary input's geometry -- including its LargestPossibleRegion -- onto
+  // the outputs. The feature-map geometry is data-dependent and is set in GenerateData()
+  // via graft; letting the default run (e.g. on a downstream UpdateOutputInformation
+  // re-propagation) would reset LargestPossibleRegion to the input image size while the
+  // buffer holds the smaller feature map, making a consumer iterate out of bounds. See
+  // the declaration in itkImageToFeaturesMap.h for the full rationale.
+}
+
 /**
  * ******************* pca_fit ***********************
  */