Scholarlink

<p>The process of obtaining a disparity map from a pair of images is called stereo matching, which is a key technology in fields such as autonomous driving and augmented reality. However, few methods consider the problem of large errors in disparity maps caused by scene depth, and mis-matching and discontinuity in disparity are more likely to occur in areas of scene depth. Therefore, this paper designs a stereo matching network based on Transmission Embedding Encoding Volume (TEEV), aiming to emphasize the role of potential depth information. The main content includes constructing a TEEV that emphasizes the role of potential depth information, while making the network pay more attention to information from deeper areas. In addition, this paper have designed a sub-pixel optimization module that can further optimize disparities within the range of (-1, 1). On the Sceneflow dataset, the average endpoint error (EPE) reached 0.42, and TEEV also achieved state-of-the-art (SOTA) results on the KITTI2012 and KITTI2015 datasets. Furthermore, in the generalization performance test of Middlebury 2014, the 2-pixel error rate was only 6.62.</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

Trans embedded encoding volume for stereo matching

  • 期刊论文

  • 发表日期：

    2025年7月7日

查看全文
  • Xiaoyang Zhao,
  • Zhuo Wang,
  • Zhongchao Deng,
  • Hongde Qin,
  • Zhongben Zhu

摘要

The process of obtaining a disparity map from a pair of images is called stereo matching, which is a key technology in fields such as autonomous driving and augmented reality. However, few methods consider the problem of large errors in disparity maps caused by scene depth, and mis-matching and discontinuity in disparity are more likely to occur in areas of scene depth. Therefore, this paper designs a stereo matching network based on Transmission Embedding Encoding Volume (TEEV), aiming to emphasize the role of potential depth information. The main content includes constructing a TEEV that emphasizes the role of potential depth information, while making the network pay more attention to information from deeper areas. In addition, this paper have designed a sub-pixel optimization module that can further optimize disparities within the range of (-1, 1). On the Sceneflow dataset, the average endpoint error (EPE) reached 0.42, and TEEV also achieved state-of-the-art (SOTA) results on the KITTI2012 and KITTI2015 datasets. Furthermore, in the generalization performance test of Middlebury 2014, the 2-pixel error rate was only 6.62.