Leveraging FINCH and K-Means for Enhanced Cluster-Based Instance Selection
Abstract
We introduce a novel instance selection method, which integrates FINCH and the K-means clustering algorithms into a unified process for improved instance selection. Initially, FINCH is employed to perform a first dataset clustering. Representatives of the resulting clusters are used as seeds for K-means to refine clustering. The samples that are closer to the centers of the tuned clusters form the set of selected instances. Our experiments show that our method outperforms established instance selection techniques. We also showcase the practical benefits of our approach by applying it to reduce the size of augmented training datasets in a case-study that involves ship detection in aerial and satellite images. The results demonstrate that our method leads to significant dataset size reduction with minimal impact on ship detection accuracy.
Cite
Text
Zotou et al. "Leveraging FINCH and K-Means for Enhanced Cluster-Based Instance Selection." European Conference on Computer Vision Workshops, 2024. doi:10.1007/978-3-031-93806-1_2Markdown
[Zotou et al. "Leveraging FINCH and K-Means for Enhanced Cluster-Based Instance Selection." European Conference on Computer Vision Workshops, 2024.](https://mlanthology.org/eccvw/2024/zotou2024eccvw-leveraging/) doi:10.1007/978-3-031-93806-1_2BibTeX
@inproceedings{zotou2024eccvw-leveraging,
title = {{Leveraging FINCH and K-Means for Enhanced Cluster-Based Instance Selection}},
author = {Zotou, Panagiota and Bacharidis, Konstantinos and Argyros, Antonis A.},
booktitle = {European Conference on Computer Vision Workshops},
year = {2024},
pages = {10-26},
doi = {10.1007/978-3-031-93806-1_2},
url = {https://mlanthology.org/eccvw/2024/zotou2024eccvw-leveraging/}
}