org.openimaj.knn.pq

Class FloatADCNearestNeighbours

java.lang.Object

org.openimaj.knn.FloatNearestNeighbours

org.openimaj.knn.pq.FloatADCNearestNeighbours

All Implemented Interfaces:

NearestNeighbours<float[],float[],IntFloatPair>

Direct Known Subclasses:

FloatSDCNearestNeighbours

@Reference(type=Article,
           author={"Jegou, Herve","Douze, Matthijs","Schmid, Cordelia"},
           title="Product Quantization for Nearest Neighbor Search",
           year="2011",
           journal="IEEE Trans. Pattern Anal. Mach. Intell.",
           pages={"117","","128"},
           url="http://dx.doi.org/10.1109/TPAMI.2010.57",
           month="January",
           number="1",
           publisher="IEEE Computer Society",
           volume="33",
           customData={"issn","0162-8828","numpages","12","doi","10.1109/TPAMI.2010.57","acmid","1916695","address","Washington, DC, USA","keywords","High-dimensional indexing, High-dimensional indexing, image indexing, very large databases, approximate search., approximate search., image indexing, very large databases"})
public class FloatADCNearestNeighbours
extends FloatNearestNeighbours

Nearest-neighbours using Asymmetric Distance Computation (ADC) on Product Quantised vectors. In ADC, only the database points are quantised. The queries themselves are not quantised. The overall distance is computed as the summed distance of each subvector of the query to each corresponding centroids of each database vector.

For efficiency, the distance of each sub-vector of a query is computed to every centroid (for the sub-vector under consideration) only once, and is then cached for the lookup during the computation of the distance to each database vector.

Author:

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

Field Summary

Fields

Modifier and Type	Field and Description
protected byte[][]	data
protected int	ndims
protected FloatProductQuantiser	pq

Constructor Summary

Constructors

Constructor and Description
FloatADCNearestNeighbours(FloatProductQuantiser pq, byte[][] pqData, int ndims) Construct the ADC with the given quantiser and pre-quantised data .
FloatADCNearestNeighbours(FloatProductQuantiser pq, float[][] dataPoints) Construct the ADC with the given quantiser and data points.

Method Summary

Methods

Modifier and Type	Method and Description
protected void	computeDistances(float[] fullQuery, BoundedPriorityQueue<IntFloatPair> queue, IntFloatPair wp)
int	numDimensions() Get the number of dimensions of each vector in the dataset
void	searchKNN(float[][] qus, int K, int[][] indices, float[][] distances) Search for the K nearest neighbours to each of the N queries, and return the indices of each nearest neighbour and their respective distances.
List<IntFloatPair>	searchKNN(float[] query, int K) Search for the K nearest neighbours to the given query and return an ordered list of pairs containing the distance and index of each neighbour.
void	searchKNN(List<float[]> qus, int K, int[][] indices, float[][] distances) Search for the K nearest neighbours to each of the N queries, and return the indices of each nearest neighbour and their respective distances.
IntFloatPair	searchNN(float[] query) Search for the nearest neighbour to the given query and return a pair containing the distance and index of that neighbour.
void	searchNN(float[][] qus, int[] indices, float[] distances) Search for the nearest neighbour to each of the N queries, and return the index of each nearest neighbour and the respective distance.
void	searchNN(List<float[]> qus, int[] indices, float[] distances) Search for the nearest neighbour to each of the N queries, and return the index of each nearest neighbour and the respective distance.
int	size() Get the size of the dataset

Methods inherited from class org.openimaj.knn.FloatNearestNeighbours

distanceFunc, distanceFunc, distanceFunc, distanceFunc

Methods inherited from class java.lang.Object

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

Field Detail

pq

protected final FloatProductQuantiser pq

ndims

protected final int ndims

data

protected final byte[][] data

Constructor Detail

FloatADCNearestNeighbours

public FloatADCNearestNeighbours(FloatProductQuantiser pq,
                         float[][] dataPoints)

Construct the ADC with the given quantiser and data points.

Parameters:

pq - the Product Quantiser

dataPoints - the data points to index

FloatADCNearestNeighbours

public FloatADCNearestNeighbours(FloatProductQuantiser pq,
                         byte[][] pqData,
                         int ndims)

Construct the ADC with the given quantiser and pre-quantised data .

Parameters:

pq - the Product Quantiser

pqData - the pre-quantised data (i.e. vectors already quantised with the given pq)

ndims - the dimensionality of the indexed data

Method Detail

searchNN

public void searchNN(float[][] qus,
            int[] indices,
            float[] distances)

Description copied from interface: NearestNeighbours

Search for the nearest neighbour to each of the N queries, and return the index of each nearest neighbour and the respective distance.

For efficiency, to use this method, you need to pre-construct the arrays for storing the results outside of the method and pass them in as arguments.

If a nearest-neighbour cannot be determined, it will have an index value of -1

Parameters:

qus - An array of N query vectors

indices - The return N-dimensional array for holding the indices of the nearest neighbour of each respective query.

distances - The return N-dimensional array for holding the distances of the nearest neighbour to each respective query.

searchKNN

public void searchKNN(float[][] qus,
             int K,
             int[][] indices,
             float[][] distances)

Description copied from interface: NearestNeighbours

Search for the K nearest neighbours to each of the N queries, and return the indices of each nearest neighbour and their respective distances.

For efficiency, to use this method, you need to pre-construct the arrays for storing the results outside of the method and pass them in as arguments.

If a k-th nearest-neighbour cannot be determined, it will have an index value of -1

Parameters:

qus - An array of N query vectors

K - the number of neighbours to find

indices - The return N*K-dimensional array for holding the indices of the K nearest neighbours of each respective query.

distances - The return N*K-dimensional array for holding the distances of the nearest neighbours of each respective query.

searchNN

public void searchNN(List<float[]> qus,
            int[] indices,
            float[] distances)

Description copied from interface: NearestNeighbours

Search for the nearest neighbour to each of the N queries, and return the index of each nearest neighbour and the respective distance.

For efficiency, to use this method, you need to pre-construct the arrays for storing the results outside of the method and pass them in as arguments.

If a nearest-neighbour cannot be determined, it will have an index value of -1

Parameters:

qus - An array of N query vectors

indices - The return N-dimensional array for holding the indices of the nearest neighbour of each respective query.

distances - The return N-dimensional array for holding the distances of the nearest neighbour to each respective query.

searchKNN

public void searchKNN(List<float[]> qus,
             int K,
             int[][] indices,
             float[][] distances)

Description copied from interface: NearestNeighbours

Search for the K nearest neighbours to each of the N queries, and return the indices of each nearest neighbour and their respective distances.

For efficiency, to use this method, you need to pre-construct the arrays for storing the results outside of the method and pass them in as arguments.

If a k-th nearest-neighbour cannot be determined, it will have an index value of -1

Parameters:

qus - An array of N query vectors

K - the number of neighbours to find

indices - The return N*K-dimensional array for holding the indices of the K nearest neighbours of each respective query.

distances - The return N*K-dimensional array for holding the distances of the nearest neighbours of each respective query.

searchKNN

public List<IntFloatPair> searchKNN(float[] query,
                           int K)

Description copied from interface: NearestNeighbours

Search for the K nearest neighbours to the given query and return an ordered list of pairs containing the distance and index of each neighbour.

If k neighbours cannot be determined, then the resultant list might have fewer than k elements.

Parameters:

query - the query vector

K - the number of neighbours to search for

Returns:

the top K nearest neighbours ordered by increasing distance

searchNN

public IntFloatPair searchNN(float[] query)

Description copied from interface: NearestNeighbours

Search for the nearest neighbour to the given query and return a pair containing the distance and index of that neighbour.

If the nearest-neighbour cannot be determined null will be returned.

Parameters:

query - the query vector

Returns:

the distance and index of the nearest neighbour

computeDistances

protected void computeDistances(float[] fullQuery,
                    BoundedPriorityQueue<IntFloatPair> queue,
                    IntFloatPair wp)

numDimensions

public int numDimensions()

Description copied from class: FloatNearestNeighbours

Get the number of dimensions of each vector in the dataset

Specified by:

numDimensions in class FloatNearestNeighbours

Returns:

the number of dimensions

size

public int size()

Description copied from interface: NearestNeighbours

Get the size of the dataset

Returns:

the dataset size