Feature Accumulators

The main work of compute_features() is performed by “accumulator” classes. There are three types of acccumulators: sp, edge, and flatedge.

  • sp features process the dense volume of labels and values.

  • edge accumulators process tables of edge pixels (and corresponding values).

  • flatedge accumulators process edges between flat superpixels. flatedge accumulators can only be used if your Rag is constructed with flat_superpixels=True.

Built-in

Base classes

You can write your own accumulator classes and pass them in an accumulator_set to compute_features(). All accumulators must inherit from one of these two base classes:

Reference

Standard Accumulators

class ilastikrag.accumulators.standard.StandardEdgeAccumulator(rag, feature_names)[source]

Accumulator for features of the edge pixels (only) between superpixels. Uses vigra’s RegionFeatureAccumulator library to compute the features.

Supported feature names:

  • standard_edge_count

  • standard_edge_sum

  • standard_edge_minimum

  • standard_edge_maximum

  • standard_edge_mean

  • standard_edge_variance

  • standard_edge_kurtosis

  • standard_edge_skewness

  • standard_edge_quantiles (short for “all edge quantiles”)

  • standard_edge_quantiles_0

  • standard_edge_quantiles_10

  • standard_edge_quantiles_25

  • standard_edge_quantiles_50

  • standard_edge_quantiles_75

  • standard_edge_quantiles_90

  • standard_edge_quantiles_100

ACCUMULATOR_TYPE = 'edge'

Accumulator type

ACCUMULATOR_ID = 'standard'

An id string for this accumulator. Must not contain an underscore (_). Must not conflict with any other accumulators of the same type (‘edge’). All feature names supported by this accumulator must begin with the prefix <id>_edge_

Methods: See BaseEdgeAccumulator

class ilastikrag.accumulators.standard.StandardSpAccumulator(rag, feature_names)[source]

Accumulator for features of the superpixels contents. Uses vigra’s RegionFeatureAccumulator library to compute the features.

Supported feature names:

  • standard_sp_count

  • standard_sp_sum

  • standard_sp_minimum

  • standard_sp_maximum

  • standard_sp_mean

  • standard_sp_variance

  • standard_sp_kurtosis

  • standard_sp_skewness

  • standard_sp_quantiles (short for “all sp quantiles”)

  • standard_sp_quantiles_0

  • standard_sp_quantiles_10

  • standard_sp_quantiles_25

  • standard_sp_quantiles_50

  • standard_sp_quantiles_75

  • standard_sp_quantiles_90

  • standard_sp_quantiles_100

  • standard_sp_regionradii (all of the below)

  • standard_sp_regionradii_0

  • standard_sp_regionradii_1

  • standard_sp_regionradii_2

  • standard_sp_regionaxes (all of the below)

  • standard_sp_regionaxes_0x

  • standard_sp_regionaxes_0y

  • standard_sp_regionaxes_0z

  • standard_sp_regionaxes_1x

  • standard_sp_regionaxes_1y

  • standard_sp_regionaxes_1z

  • standard_sp_regionaxes_2x

  • standard_sp_regionaxes_2y

  • standard_sp_regionaxes_2z

All input feature names result in two output columns, for the _sum and _difference between the two superpixels adjacent to the edge.

As a special case, the output columns for the sp_count feature are reduced via cube-root (or square-root), as specified in the multicut paper.

ACCUMULATOR_TYPE = 'sp'

Accumulator type

ACCUMULATOR_ID = 'standard'

An id string for this accumulator. Must not contain an underscore (_). Must not conflict with any other accumulators of the same type (‘sp’). All feature names supported by this accumulator must begin with the prefix <id>_sp_

Methods: See BaseSpAccumulator

class ilastikrag.accumulators.standard.StandardFlatEdgeAccumulator(rag, feature_names)[source]

Accumulator for basic features of ‘flat edges’, i.e. the edges in the Z-direction when Rag.flat_superpixels == True.

Supported feature names:

  • standard_flatedge_count

  • standard_flatedge_sum

  • standard_flatedge_minimum

  • standard_flatedge_maximum

  • standard_flatedge_mean

  • standard_flatedge_variance

  • standard_flatedge_kurtosis

  • standard_flatedge_skewness

  • standard_flatedge_quantiles (short for “all sp quantiles”)

  • standard_flatedge_quantiles_0

  • standard_flatedge_quantiles_10

  • standard_flatedge_quantiles_25

  • standard_flatedge_quantiles_50

  • standard_flatedge_quantiles_75

  • standard_flatedge_quantiles_90

  • standard_flatedge_quantiles_100

  • standard_flatedge_regionradii (both of the below)

  • standard_flatedge_regionradii_0

  • standard_flatedge_regionradii_1

  • standard_flatedge_regionaxes (all of the below)

  • standard_flatedge_regionaxes_0x

  • standard_flatedge_regionaxes_0y

  • standard_flatedge_regionaxes_1x

  • standard_flatedge_regionaxes_1y

  • standard_flatedge_correlation

ACCUMULATOR_TYPE = 'flatedge'

Accumulator type

ACCUMULATOR_ID = 'standard'

An id string for this accumulator. Must not contain an underscore (_). Must not conflict with any other accumulators of the same type (‘sp’). All feature names supported by this accumulator must begin with the prefix <id>_sp_

Methods: See BaseFlatEdgeAccumulator

EdgeRegion Accumulator

class ilastikrag.accumulators.edgeregion.EdgeRegionEdgeAccumulator(rag, feature_names)[source]

Accumulator for computing region axes and region radii over edge coordinates. (The StandardEdgeAccumulator class does not provide region features.)

We don’t use vigra’s RegionFeatureAccumulators because we only have access to the sparse lists of edge pixels (along each axis). Instead, we manually compute the region axes/radii directly from the edge coordinate columns.

Supported feature names:

  • edgeregion_edge_area (radii_0 * radii_1)

  • edgeregion_edge_volume (radii_0 * radii_1 * radii_2)

  • edgeregion_edge_regionradii (all of the below)

  • edgeregion_edge_regionradii_0

  • edgeregion_edge_regionradii_1

  • edgeregion_edge_regionradii_2

  • edgeregion_edge_regionaxes (all of the below)

  • edgeregion_edge_regionaxes_0x

  • edgeregion_edge_regionaxes_0y

  • edgeregion_edge_regionaxes_0z

  • edgeregion_edge_regionaxes_1x

  • edgeregion_edge_regionaxes_1y

  • edgeregion_edge_regionaxes_1z

  • edgeregion_edge_regionaxes_2x

  • edgeregion_edge_regionaxes_2y

  • edgeregion_edge_regionaxes_2z

ACCUMULATOR_TYPE = 'edge'

Accumulator type

ACCUMULATOR_ID = 'edgeregion'

An id string for this accumulator. Must not contain an underscore (_). Must not conflict with any other accumulators of the same type (‘edge’). All feature names supported by this accumulator must begin with the prefix <id>_edge_

Methods: See BaseEdgeAccumulator

Similarity FlatEdge Accumulator

class ilastikrag.accumulators.similarity.SimilarityFlatEdgeAccumulator(rag, feature_names)[source]

Accumulator for computing various similarity metrics between pixel values adjacent to ‘flat edges’, i.e. the edges in the Z-direction when Rag.flat_superpixels == True.

Supported feature names:

  • similarity_flatedge_correlation

ACCUMULATOR_TYPE = 'flatedge'

Accumulator type

ACCUMULATOR_ID = 'similarity'

An id string for this accumulator. Must not contain an underscore (_). Must not conflict with any other accumulators of the same type (‘sp’). All feature names supported by this accumulator must begin with the prefix <id>_sp_

Methods: See BaseFlatEdgeAccumulator

Base Accumulators

class ilastikrag.accumulators.base.BaseEdgeAccumulator(rag, feature_names)[source]

Base class for all edge accumulators, i.e. accumulators that compute features from the edge values between superpixels (not superpixel contents).

ACCUMULATOR_TYPE = 'edge'

Accumulator type

ACCUMULATOR_ID = ''

An id string for this accumulator. Must not contain an underscore (_). Must not conflict with any other accumulators of the same type (‘edge’). All feature names supported by this accumulator must begin with the prefix <id>_edge_

__init__(rag, feature_names)[source]
Parameters:

  • rag – The rag.

  • feature_names – A list of feature names to compute with this accumulator.

cleanup()[source]

Called by the Rag to indicate that processing has completed, and the accumulator should discard all cached data and intermediate results. Subclasses must reimplement this function.

ingest_edges(rag, edge_values)[source]

Ingests the given edge values using the given Rag.

Parameters:

  • rag – Rag

  • edge_values – OrderedDict of 1D ndarray. Each ndarray edge_values[k] is in the same order as rag.dense_edges[k]

append_edge_features_to_df(edge_df)[source]

Called by the Rag after ingest_edges().

Merges the features of all ingested edges into a final set of edge feature columns, and appends those columns to the given pandas.DataFrame object.

classmethod supported_features(rag)[source]

Returns the list of feature names that can be computed for the given Rag.

class ilastikrag.accumulators.base.BaseSpAccumulator(rag, feature_names)[source]

Base class for all superpixel accumulators, i.e. accumulators that compute features from the contents of superpixels (not their edges).

ACCUMULATOR_TYPE = 'sp'

Accumulator type

ACCUMULATOR_ID = ''

An id string for this accumulator. Must not contain an underscore (_). Must not conflict with any other accumulators of the same type (‘sp’). All feature names supported by this accumulator must begin with the prefix <id>_sp_

__init__(rag, feature_names)[source]
Parameters:

  • rag – The rag.

  • feature_names – A list of feature names to compute with this accumulator.

cleanup()[source]

Called by the Rag to indicate that processing has completed, and the accumulator should discard all cached data and intermediate results. Subclasses must reimplement this function.

ingest_values(rag, value_img)[source]

Ingest the given (single-channel) pixel values, using the superpixels stored in rag.label_img.

Parameters:

  • rag – Rag

  • value_img – VigraArray, same shape as rag.label_img

append_edge_features_to_df(edge_df)[source]

Called by the Rag after ingest_values().

Merges the features of ingested data into a final set of edge feature columns, and appends those columns to the given pandas.DataFrame object.

This involves converting pairs superpixel features into edge features, typically by taking the sum and/or difference between the features of each superpixel in an adjacent pair.

classmethod supported_features(rag)[source]

Returns the list of feature names that can be computed for the given Rag.

class ilastikrag.accumulators.base.BaseFlatEdgeAccumulator(rag, feature_names)[source]

Base class for all flatedge accumulators, i.e. accumulators that compute features over the edges in the Z direction when Rag.flat_superpixels == True

ACCUMULATOR_TYPE = 'flatedge'

Accumulator type

ACCUMULATOR_ID = ''

An id string for this accumulator. Must not contain an underscore (_). Must not conflict with any other accumulators of the same type (‘sp’). All feature names supported by this accumulator must begin with the prefix <id>_sp_

__init__(rag, feature_names)[source]
Parameters:

  • rag – The rag.

  • feature_names – A list of feature names to compute with this accumulator.

cleanup()[source]

Called by the Rag to indicate that processing has completed, and the accumulator should discard all cached data and intermediate results. Subclasses must reimplement this function.

ingest_values(rag, value_img)[source]

Ingest the given (single-channel) pixel values, using the (flat) superpixels stored in rag.label_img.

Parameters:

  • rag – Rag

  • value_img – VigraArray, same shape as rag.label_img

append_edge_features_to_df(edge_df)[source]

Called by the Rag after ingest_values().

Merges the features of all ingested edges into a final set of edge feature columns, and appends those columns to the given pandas.DataFrame object.

classmethod supported_features(rag)[source]

Returns the list of feature names that can be computed for the given Rag.