Scholarlink

<p>The performance evaluation of a change detection algorithm aims to assess its effectiveness compared to state-of-the-art methods. This process involves running the algorithm to detect changes in video frames and comparing its outputs against a ground truth. An effective evaluation requires a diverse set of videos with varying difficulty levels, ensuring that each video contributes unique information about the algorithm’s performance. However, existing datasets often contain redundant samples with similar difficulty levels, which can lead to inefficiencies. In this work, we propose a method to remove redundant videos while preserving the evaluation potential of the original dataset. We tested our approach on the CDNet 2014 dataset, demonstrating that it reduces the dataset size by approximately 45% without compromising the validity of algorithm assessments.</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

Optimizing Change Detection Evaluation: Reducing Dataset Redundancy While Preserving Performance Assessment

  • 期刊论文

  • 发表日期：

    2025年12月15日

查看全文
  • Silvio R. R. Sanches,
  • Priscila T. M. Saito,
  • Pedro H. Bugatti,
  • Elton Custódio Junior,
  • Valdinei Freire

摘要

The performance evaluation of a change detection algorithm aims to assess its effectiveness compared to state-of-the-art methods. This process involves running the algorithm to detect changes in video frames and comparing its outputs against a ground truth. An effective evaluation requires a diverse set of videos with varying difficulty levels, ensuring that each video contributes unique information about the algorithm’s performance. However, existing datasets often contain redundant samples with similar difficulty levels, which can lead to inefficiencies. In this work, we propose a method to remove redundant videos while preserving the evaluation potential of the original dataset. We tested our approach on the CDNet 2014 dataset, demonstrating that it reduces the dataset size by approximately 45% without compromising the validity of algorithm assessments.