Scholarlink

Abstract <p>A new data selection criterion for active learning of object detectors based on deep neural networks with feature pyramid network (FPN) architecture is presented. The key innovation of the method is the consideration of the local representation of embeddings of objects of interest at different levels of the pyramidal network. To do this, the RoiAlign region of interest alignment operation is used, which avoids distortions due to scaling and rounding. Local embeddings corresponding to labeled objects of interest are processed using principal component analysis and used to select new data based on cosine distance. The effectiveness of the proposed criterion is confirmed experimentally by iteratively training the YOLOv8 detector on the COCO2017 and Pascal VOC-2017 datasets. Comparative analysis with traditional confidence selection methods showed a stable improvement in detection quality measures. The method can be applied in the problems of filtering pseudolabeled data sets, selecting synthetic data, as well as in related problems such as searching for similar objects, hard negative mining, and evaluating the effectiveness of augmentations.</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

Extracting Embeddings for Active Learning Corresponding to Regions of Interest in an Image

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年9月1日

查看全文
  • V. A. Sokolov,
  • I. A. Matveev,
  • L. M. Teplyakov

摘要

Abstract

A new data selection criterion for active learning of object detectors based on deep neural networks with feature pyramid network (FPN) architecture is presented. The key innovation of the method is the consideration of the local representation of embeddings of objects of interest at different levels of the pyramidal network. To do this, the RoiAlign region of interest alignment operation is used, which avoids distortions due to scaling and rounding. Local embeddings corresponding to labeled objects of interest are processed using principal component analysis and used to select new data based on cosine distance. The effectiveness of the proposed criterion is confirmed experimentally by iteratively training the YOLOv8 detector on the COCO2017 and Pascal VOC-2017 datasets. Comparative analysis with traditional confidence selection methods showed a stable improvement in detection quality measures. The method can be applied in the problems of filtering pseudolabeled data sets, selecting synthetic data, as well as in related problems such as searching for similar objects, hard negative mining, and evaluating the effectiveness of augmentations.