org.openimaj.image.indexing.vlad

Class VLADIndexerDataBuilder

java.lang.Object

org.openimaj.image.indexing.vlad.VLADIndexerDataBuilder

public class VLADIndexerDataBuilder
extends Object

Class for learning the data required to efficiently index images using VLAD with PCA and product quantisers.

Author:

Jonathon Hare (jsh2@ecs.soton.ac.uk)

Nested Class Summary

Nested Classes

Modifier and Type	Class and Description
static class	VLADIndexerDataBuilder.StandardPostProcesses Feature post-processing options

Constructor Summary

Constructors

Constructor and Description

VLADIndexerDataBuilder(LocalFeatureExtractor<LocalFeature<?,?>,MBFImage> extractor, List<File> localFeatures, boolean normalise, int numVladCentroids, int numIterations, int numPcaDims, int numPqIterations, int numPqAssigners, float sampleProp, float pcaSampleProp, Function<List<? extends LocalFeature<?,?>>,List<FloatLocalFeatureAdaptor<?>>> postProcess) Construct a VLADIndexerDataBuilder with the given parameters

Method Summary

Modifier and Type	Method and Description
VLADIndexerData	buildIndexerData() Build the VLADIndexerData using the information provided at construction time.
VLAD<float[]>	buildVLAD() Build a VLAD using the information provided at construction time.

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Constructor Detail

VLADIndexerDataBuilder

public VLADIndexerDataBuilder(LocalFeatureExtractor<LocalFeature<?,?>,MBFImage> extractor,
                              List<File> localFeatures,
                              boolean normalise,
                              int numVladCentroids,
                              int numIterations,
                              int numPcaDims,
                              int numPqIterations,
                              int numPqAssigners,
                              float sampleProp,
                              float pcaSampleProp,
                              Function<List<? extends LocalFeature<?,?>>,List<FloatLocalFeatureAdaptor<?>>> postProcess)

Construct a VLADIndexerDataBuilder with the given parameters

Parameters:

extractor - the local feature extractor used to generate the input features

localFeatures - a list of file locations of the files containing the input local features (one per image)

normalise - should the resultant VLAD features be l2 normalised?

numVladCentroids - the number of centroids for VLAD (~64)

numIterations - the number of clustering iterations (~100)

numPcaDims - the number of dimensions to project down to using PCA (~128 for normal SIFT)

numPqIterations - the number of iterations for clustering the product quantisers (~100)

numPqAssigners - the number of product quantiser assigners (~16)

sampleProp - the proportion of features to sample for the clustering the VLAD centroids

pcaSampleProp - the proportion of images to sample for computing the PCA basis

postProcess - the post-processing to apply to the raw features before input to VLAD

Method Detail

buildIndexerData

public VLADIndexerData buildIndexerData()
                                 throws IOException

Build the VLADIndexerData using the information provided at construction time. The following steps are taken:

1. A sample of the features is loaded
2. The sample is clustered using k-means
3. VLAD representations are then built for all the input images
4. PCA is performed on the VLAD features
5. Whitening is applied to the PCA basis
6. The VLAD features are projected by the basis
7. Product quantisers are learned
8. The final VLADIndexerData object is created

Returns:

a newly learned VLADIndexerData object

Throws:

IOException

buildVLAD

public VLAD<float[]> buildVLAD()

Build a VLAD using the information provided at construction time. The following steps are taken:

1. A sample of the features is loaded
2. The sample is clustered using k-means

Returns:

the VLAD