Scholarlink

In the ICPR 2024 Multi-Modal Visual Pattern Recognition Challenge-Track 1 (Tracking), we proposed a multi-modal tracking method gdi-SAM2, based on the SAM2 algorithm, which significantly improves the accuracy and robustness of target tracking by fusing RGB, depth, and infrared data. Our method first converts the bounding box (bbox) annotations in the fused dataset into pixel-level mask images, then inputs these mask images into the SAM2 tracking model. The tracking results are outputted in the form of mask images, which are then converted back to bbox for performance evaluation. Our solution achieved second place in the competition with an AUC of 0.869, demonstrating the potential of multi-modal tracking technology in handling complex scenarios.

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

Tracking Model gdi-SAM2 for RGB, Depth, and Infrared Data

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Xinglin Xie,
  • Kehuan Song,
  • Kefan Chen

摘要

In the ICPR 2024 Multi-Modal Visual Pattern Recognition Challenge-Track 1 (Tracking), we proposed a multi-modal tracking method gdi-SAM2, based on the SAM2 algorithm, which significantly improves the accuracy and robustness of target tracking by fusing RGB, depth, and infrared data. Our method first converts the bounding box (bbox) annotations in the fused dataset into pixel-level mask images, then inputs these mask images into the SAM2 tracking model. The tracking results are outputted in the form of mask images, which are then converted back to bbox for performance evaluation. Our solution achieved second place in the competition with an AUC of 0.869, demonstrating the potential of multi-modal tracking technology in handling complex scenarios.