Presentation is loading. Please wait.

Presentation is loading. Please wait.

Fast Nearest Neighbor Search in the Hamming Space

Similar presentations


Presentation on theme: "Fast Nearest Neighbor Search in the Hamming Space"— Presentation transcript:

1 Fast Nearest Neighbor Search in the Hamming Space
OS3: Multimedia Machine Learning Fast Nearest Neighbor Search in the Hamming Space Zhansheng Jiang, Lingxi Xie, Xiaotie Deng, Weiwei Xu, and Jingdong Wang

2 Nearest Neighbor Search
Nearest Neighbor (NN) Search An optimization problem for finding the closest points for a particular query point from a set of reference data points. Application Large scale similar image search

3 Nearest Neighbor Search
Search in the Hamming Space Multi-Index Hashing FLANN Fast Neighborhood Graph Search Using Cartesian Concatenation

4 Data Structure Bridge Vectors Augmented Neighborhood Graph
Similar to the concept of cluster centers in Product Quantization Computed by k-means clustering for binary vectors Minimize the average distance of each vector to its nearest bridge vector Augmented Neighborhood Graph Bridge graph – connect bridge vectors and their nearest reference vectors Neighborhood graph – connect each reference vector to its nearest reference vectors

5 Compute Bridge Vectors
Defined as K-means Clustering for Binary Vectors Initialize centers randomly by the subvectors of reference vectors assignment step: find the nearest center for the subvector of each reference vector update step: optimize the centers for each cluster (by voting in each dimension)

6 Construct Augmented Neighborhood Graph
Bridge Graph Find t nearest bridge vectors for each reference vector (by Multi-Sequence algorithm) Keep b nearest reference vectors for each bridge vector Neighborhood Graph Find k nearest reference vectors for each reference vector Approximate NN search method for large scale dataset

7

8 Reference vector: yellow
Build 2015 12/9/2018 3:26 PM A X B F C Y I E D Z G H Query vector: white Bridge vector: red Reference vector: yellow © 2015 Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.

9 Push X PQ = {X} 5NN = {} A X B F C I Y E D Z G H Build 2015
12/9/2018 3:26 PM A X B F C Y I E D Z G H Push X PQ = {X} 5NN = {} © 2015 Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.

10 Pop X and push ABCEF and Y PQ = {ECFBYA} 5NN = {}
Build 2015 12/9/2018 3:26 PM A X B F C Y I E D Z G H Pop X and push ABCEF and Y PQ = {ECFBYA} 5NN = {} © 2015 Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.

11 Pop E and Push D PQ = {CDFBYA} 5NN = {E} A X B F C I Y E D Z G H
Build 2015 12/9/2018 3:26 PM A X B F C Y I E D Z G H Pop E and Push D PQ = {CDFBYA} 5NN = {E} © 2015 Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.

12 Pop C PQ = {DFBYA} 5NN = {EC} A X B F C I Y E D Z G H Build 2015
12/9/2018 3:26 PM A X B F C Y I E D Z G H Pop C PQ = {DFBYA} 5NN = {EC} © 2015 Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.

13 Pop D and push G PQ = {FGBYA} 5NN = {ECD} A X B F C I Y E D Z G H
Build 2015 12/9/2018 3:26 PM A X B F C Y I E D Z G H Pop D and push G PQ = {FGBYA} 5NN = {ECD} © 2015 Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.

14 Pop F PQ = {GBYA} 5NN = {ECDF} A X B F C I Y E D Z G H Build 2015
12/9/2018 3:26 PM A X B F C Y I E D Z G H Pop F PQ = {GBYA} 5NN = {ECDF} © 2015 Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.

15 Pop G and push H PQ = {BHYA} 5NN = {ECDFG} A X B F C I Y E D Z G H
Build 2015 12/9/2018 3:26 PM A X B F C Y I E D Z G H Pop G and push H PQ = {BHYA} 5NN = {ECDFG} © 2015 Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.

16

17

18

19

20 Build 2015 12/9/2018 3:26 PM © 2015 Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.


Download ppt "Fast Nearest Neighbor Search in the Hamming Space"

Similar presentations


Ads by Google