Dataset Distillation with Foreground Extraction

カテゴリ：部門大会

論文No：PS6-10

グループ名：【C】2025年電気学会電子・情報・システム部門大会

発行日：2025/8/20

タイトル(英語)：Dataset Distillation with Foreground Extraction

著者名：Cai Wenqi（富山大学）,Zou Yawen（富山大学）,Zhang Chao（富山大学）

著者名(英語)： Wenqi Cai (University of Toyama),Yawen Zou (University of Toyama),Chao Zhang (University of Toyama)

キーワード：データセット蒸留,前景抽出,セマンティックな忠実性セマンティックな忠実性,Dataset distillation,Foreground Extraction,semantic fidelity

要約(日本語)：Dataset distillation compresses large datasets into compact synthetic subsets for efficient training. However, many recent approaches prioritize accuracy while overlooking the semantic and visual fidelity of generated samples. Synthesized images often lack clear foreground objects, achieving high accuracy by overfitting to background cues. To address this, we propose DDFE (Dataset Distillation with Foreground Extraction), which preserves class-relevant foregrounds to promote meaningful feature learning. While DDFE may slightly reduce accuracy, it improves visual quality and semantic alignment, exposing the background bias in prior methods and underscoring the need to assess both fidelity and performance in distillation.

本誌掲載ページ：1648-1652p

原稿種別：英語

PDFファイルサイズ：454Kバイト