Scholarlink

<p>This paper introduces an evaluation pipeline for transferability estimation in fine-grained visual recognition tasks, with a focus on hand gesture recognition. The proposed framework provides a resource-efficient and interpretable tool for validating model suitability to downstream tasks without exhaustive fine-tuning. By analyzing the representation spaces of both convolutional and transformer-based vision models through dimensionality reduction and clustering, our pipeline incorporates a similarity-aware scoring technique that explicitly considers both intra-class variability and inter-class similarity. In experiments, our framework consistently ranked the vision transformer with advanced pre-training highest, while assigning lower importance to the standard vision transformer and convolutional models. These results indicate that the proposed approach can effectively highlight models best suited for fine-grained recognition. We provide the source code and experimental setup at <a href="https://github.com/ADA-SITE-JML/gesture_face_features">https://github.com/ADA-SITE-JML/gesture_face_features</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

Representation space evaluation of feature extractors for model transferability in fine-grained recognition

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月20日

查看全文
  • Ismayil Shahaliyev,
  • Jamaladdin Hasanov

摘要

This paper introduces an evaluation pipeline for transferability estimation in fine-grained visual recognition tasks, with a focus on hand gesture recognition. The proposed framework provides a resource-efficient and interpretable tool for validating model suitability to downstream tasks without exhaustive fine-tuning. By analyzing the representation spaces of both convolutional and transformer-based vision models through dimensionality reduction and clustering, our pipeline incorporates a similarity-aware scoring technique that explicitly considers both intra-class variability and inter-class similarity. In experiments, our framework consistently ranked the vision transformer with advanced pre-training highest, while assigning lower importance to the standard vision transformer and convolutional models. These results indicate that the proposed approach can effectively highlight models best suited for fine-grained recognition. We provide the source code and experimental setup at https://github.com/ADA-SITE-JML/gesture_face_features.