Scholarlink

<p>Local feature extraction, as the core component of feature matching, directly determines matching accuracy. To address the weak representational ability in low-texture regions, the lack of long-range dependencies, and the loss of structural details in local features, this paper proposes a local feature extraction method based on feature aggregation, mixed convolution and attention mechanisms, built upon the accelerated feature extraction network. First, a feature aggregation module is introduced in the feature extraction stage to enhance the representation of local features in low-texture regions. A local detail estimation branch and an efficient approximate attention branch collaboratively capture local detail information and non-local structural information, respectively. Second, to address the lack of long-range dependencies in local features, a mixed convolution–attention module is introduced in the feature fusion stage. Through the parallel extraction of the convolutional branch and the self-attention branch, the module enables local detail capture and global dependency modeling. Finally, a dynamic sampling strategy is adopted in the upsampling stage to alleviate structural detail loss caused by fixed bilinear interpolation. By adaptively adjusting the sampling positions, cross-scale feature alignment and accurate reconstruction of structural details are achieved. Experimental results demonstrate that, on the MegaDepth dataset, the proposed method achieves AUC@5°, AUC@10°, and AUC@20° of 46.4%, 59.6%, and 70.7%, representing improvements of 6.4%, 5.1%, and 4.3% over XFeat in the downstream camera pose estimation task.</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

Local feature extraction method based on feature aggregation, mixed convolution and attention

  • 期刊论文

  • 发表日期：

    2026年4月21日

查看全文
  • Lunming Qin,
  • Mengqian Quan,
  • Haoyang Cui,
  • Liang Xue,
  • Houqin Bian,
  • Xi Wang

摘要

Local feature extraction, as the core component of feature matching, directly determines matching accuracy. To address the weak representational ability in low-texture regions, the lack of long-range dependencies, and the loss of structural details in local features, this paper proposes a local feature extraction method based on feature aggregation, mixed convolution and attention mechanisms, built upon the accelerated feature extraction network. First, a feature aggregation module is introduced in the feature extraction stage to enhance the representation of local features in low-texture regions. A local detail estimation branch and an efficient approximate attention branch collaboratively capture local detail information and non-local structural information, respectively. Second, to address the lack of long-range dependencies in local features, a mixed convolution–attention module is introduced in the feature fusion stage. Through the parallel extraction of the convolutional branch and the self-attention branch, the module enables local detail capture and global dependency modeling. Finally, a dynamic sampling strategy is adopted in the upsampling stage to alleviate structural detail loss caused by fixed bilinear interpolation. By adaptively adjusting the sampling positions, cross-scale feature alignment and accurate reconstruction of structural details are achieved. Experimental results demonstrate that, on the MegaDepth dataset, the proposed method achieves AUC@5°, AUC@10°, and AUC@20° of 46.4%, 59.6%, and 70.7%, representing improvements of 6.4%, 5.1%, and 4.3% over XFeat in the downstream camera pose estimation task.