Graph-based visual instance mining with geometric matching and nearest candidates selection

The goal of this work is to cluster frequently appearing visual instances from a collection of images, called visual instance mining. State-of-the-art approaches employ clustering on the matching graph. Although these graph-based approaches are quite effective, their performances are still modest. I...

Full description

Saved in:
Bibliographic Details
Published in2017 9th International Conference on Knowledge and Systems Engineering (KSE) pp. 263 - 268
Main Authors Ngoc-Bao Nguyen, Nguyen, Khang M. T. T., Cuong Mai Van, Duy-Dinh Le
Format Conference Proceeding
LanguageEnglish
Published IEEE 01.10.2017
Subjects
Burstiness
Clustering algorithms
Clustering methods
Correlation
Feature extraction
Geometric Matching
Graph Clustering
Hamming distance
Hamming Embedding
Information technology
Instance Mining
Visualization
Online AccessGet full text

Cover

More Information
Summary:The goal of this work is to cluster frequently appearing visual instances from a collection of images, called visual instance mining. State-of-the-art approaches employ clustering on the matching graph. Although these graph-based approaches are quite effective, their performances are still modest. In this paper, we propose two methods to further improve the accuracy of graph-based approaches. Firstly, we improve graph construction by integrating geometric constraints. Particularly, we exploit Hamming Embedding code, burstiness removal, and geometric consistency in our framework to remove false connections. Secondly, unlike most of the existing methods which cluster images similar to an entry point, we cluster them by prioritizing the correlation between the candidates and members in the cluster. We enlarge the cluster by adding the nearest candidate at each re-expanding. Our method outperforms the state-of-the-art methods on benchmark datasets. Especially, our results are 21.2% and 5.5% higher than the latest results, conducted respectively in MQA dataset and PartialDup dataset.
DOI:10.1109/KSE.2017.8119469