Scholarlink

<p>As a fundamental building block in computer vision, edges, categorized by the discontinuity in <i>surface-Reflectance</i>, <i>Illumination</i>, <i>surface-Normal</i> or <i>Depth</i>, are critical for scene understanding. Despite significant progress in detecting generic or individual edge types, a holistic approach to simultaneously identifying all four edge types presents a unique challenge. In this paper, we propose RINDNet++, a novel neural network solution to address this challenge by detecting all four edge types. RINDNet++ is designed with a three-stage process that effectively leverages the distinct attributes of each edge type and their relationship. In Stage I, a common backbone extracts hierarchical features for all edges. Stage II then tailors these features for each edge type using specialized decoders. Stage III predicts the initial detection results with independent decision heads based on the enhanced features from the preceding stages. Additionally, RINDNet++ incorporates an attention module that refines edge detection by highlighting inter-type relationships, leading to enhanced edge maps. To enhance rigorous training and evaluation, we introduce the first benchmark, BSDS-RIND, incorporating annotations for all four edge types and supporting both single-scale and multi-scale testing. With the integration of state-of-the-art edge detection methods, BSDS-RIND establishes a robust framework for performance evaluation. Extensive experiments show that our proposed RINDNet++ yields promising results in comparison with the state-of-the-art approaches. Moreover, RINDNet++ excels in detecting generic edges and enhances performance in downstream applications such as shadow detection and depth estimation.</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

RINDNet++: Edge Detection for Discontinuity in Reflectance, Illumination, Normal, and Depth

  • 期刊论文

  • 发表日期：

    2025年7月31日

查看全文
  • Mengyang Pu,
  • Yaping Huang,
  • Qingji Guan,
  • Zhihao Liu,
  • Haibin Ling

摘要

As a fundamental building block in computer vision, edges, categorized by the discontinuity in surface-Reflectance, Illumination, surface-Normal or Depth, are critical for scene understanding. Despite significant progress in detecting generic or individual edge types, a holistic approach to simultaneously identifying all four edge types presents a unique challenge. In this paper, we propose RINDNet++, a novel neural network solution to address this challenge by detecting all four edge types. RINDNet++ is designed with a three-stage process that effectively leverages the distinct attributes of each edge type and their relationship. In Stage I, a common backbone extracts hierarchical features for all edges. Stage II then tailors these features for each edge type using specialized decoders. Stage III predicts the initial detection results with independent decision heads based on the enhanced features from the preceding stages. Additionally, RINDNet++ incorporates an attention module that refines edge detection by highlighting inter-type relationships, leading to enhanced edge maps. To enhance rigorous training and evaluation, we introduce the first benchmark, BSDS-RIND, incorporating annotations for all four edge types and supporting both single-scale and multi-scale testing. With the integration of state-of-the-art edge detection methods, BSDS-RIND establishes a robust framework for performance evaluation. Extensive experiments show that our proposed RINDNet++ yields promising results in comparison with the state-of-the-art approaches. Moreover, RINDNet++ excels in detecting generic edges and enhances performance in downstream applications such as shadow detection and depth estimation.