Scholarlink

<p>Video object detection (VOD) is a challenging task, and image object detectors are difficult to detect degradation phenomena in certain video frames. However, existing research on VOD mostly trades high computational costs for accuracy, making it difficult to achieve a balance between accuracy and speed. This work proposes an optimized Real-Time Detection Transformer (RT-DETR) model for VOD that introduces a decoupled Feature Aggregation Module (FAM) to separately refine the localization and classification detection heads. This method only requires a minimal increase in the number of parameters to achieve significant improvements in accuracy. Specifically, we insert FAM before the localization detection head and classification detection head, and first freeze all parameters of the feature extractor and classification detection head to train only the parameters of the localization detection head to obtain more accurate localization results. Then, we freeze all parameters of the feature extractor and localization detection head to train only the parameters of the classification detection head to improve the final detection accuracy. We have conducted a large number of ablation experiments to verify the effectiveness of the method. Without using any post-processing methods, we achieved 90.0% mAP on the ImageNet-VID dataset, with only 77.9&#xa0;M parameters and an average inference speed of 14.1ms.</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

Optimized RT-DETR for accurate and efficient video object detection via decoupled feature aggregation

  • 期刊论文

  • 发表日期：

    2025年2月12日

查看全文
  • Hao Chen,
  • Wu Huang,
  • Tao Zhang

摘要

Video object detection (VOD) is a challenging task, and image object detectors are difficult to detect degradation phenomena in certain video frames. However, existing research on VOD mostly trades high computational costs for accuracy, making it difficult to achieve a balance between accuracy and speed. This work proposes an optimized Real-Time Detection Transformer (RT-DETR) model for VOD that introduces a decoupled Feature Aggregation Module (FAM) to separately refine the localization and classification detection heads. This method only requires a minimal increase in the number of parameters to achieve significant improvements in accuracy. Specifically, we insert FAM before the localization detection head and classification detection head, and first freeze all parameters of the feature extractor and classification detection head to train only the parameters of the localization detection head to obtain more accurate localization results. Then, we freeze all parameters of the feature extractor and localization detection head to train only the parameters of the classification detection head to improve the final detection accuracy. We have conducted a large number of ablation experiments to verify the effectiveness of the method. Without using any post-processing methods, we achieved 90.0% mAP on the ImageNet-VID dataset, with only 77.9 M parameters and an average inference speed of 14.1ms.