Scholarlink

<p>Real-world datasets often exhibit long-tail distributions, compromising the generalization and fairness of learning-based models. This issue is particularly pronounced in Image Aesthetics Assessment (IAA) tasks, where such imbalance is difficult to mitigate due to the severe long-tail distribution across both aesthetic labels and image scenes, as well as the ambiguity and uncontrollability of aesthetics. To address these issues, we propose an Enhancer against Long-Tail for Aesthetics-oriented models (ELTA) from two levels. 1) At the label level, ELTA employs a dedicated mixup technique to augment the feature representation of minority data, while aligning features and their intrinsic aesthetic labels through a similarity consistency approach, effectively alleviating the distribution mismatch and label ambiguity. Furthermore, ELTA refines the model’s output distribution to improve the quality of pseudo-labels. 2) At the scene level, ELTA automatically identifies long-tail scenes in the dataset, and then adopts an aesthetics-guided Text-to-Image (AG-T2I) model to generate new data with controlled scene content, image style and aesthetic quality, thus balancing scene distribution while preserving label distribution. Experiments on four representative datasets (AVA, AADB, TAD66K, and PARA) demonstrate that our proposed ELTA achieves state-of-the-art performance by effectively mitigating the long-tail issue. To foster further research, we developed comprehensive benchmarks that evaluate 16 methods and released a supplementary dataset of over 50,000 images to alleviate imbalances in the existing dataset. To make these advancements fully accessible, we also provide a plug-and-play version of ELTA for easy integration with existing IAA methods, and a complete toolkit for long-tail analysis, automated data generation, and annotation. All resources will be available at <a href="https://github.com/woshidandan/Long-Tail-image-aesthetics-and-quality-assessment/blob/main/ELTA">https://github.com/woshidandan/Long-Tail-image-aesthetics-and-quality-assessment/blob/main/ELTA</a></p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

ELTA 2.0: Rethinking Long-Tail for Image Aesthetics Assessment

  • 期刊论文

  • 发表日期：

    2026年6月25日

查看全文
  • Shuai He,
  • Yuxin Chen,
  • Limin Liu,
  • Anlong Ming,
  • Huadong Ma

摘要

Real-world datasets often exhibit long-tail distributions, compromising the generalization and fairness of learning-based models. This issue is particularly pronounced in Image Aesthetics Assessment (IAA) tasks, where such imbalance is difficult to mitigate due to the severe long-tail distribution across both aesthetic labels and image scenes, as well as the ambiguity and uncontrollability of aesthetics. To address these issues, we propose an Enhancer against Long-Tail for Aesthetics-oriented models (ELTA) from two levels. 1) At the label level, ELTA employs a dedicated mixup technique to augment the feature representation of minority data, while aligning features and their intrinsic aesthetic labels through a similarity consistency approach, effectively alleviating the distribution mismatch and label ambiguity. Furthermore, ELTA refines the model’s output distribution to improve the quality of pseudo-labels. 2) At the scene level, ELTA automatically identifies long-tail scenes in the dataset, and then adopts an aesthetics-guided Text-to-Image (AG-T2I) model to generate new data with controlled scene content, image style and aesthetic quality, thus balancing scene distribution while preserving label distribution. Experiments on four representative datasets (AVA, AADB, TAD66K, and PARA) demonstrate that our proposed ELTA achieves state-of-the-art performance by effectively mitigating the long-tail issue. To foster further research, we developed comprehensive benchmarks that evaluate 16 methods and released a supplementary dataset of over 50,000 images to alleviate imbalances in the existing dataset. To make these advancements fully accessible, we also provide a plug-and-play version of ELTA for easy integration with existing IAA methods, and a complete toolkit for long-tail analysis, automated data generation, and annotation. All resources will be available at https://github.com/woshidandan/Long-Tail-image-aesthetics-and-quality-assessment/blob/main/ELTA