Scholarlink

<p>LiDAR semantic segmentation degrades severely in adverse weather due to heterogeneous, modality-dependent domain shifts in geometric structure and reflectance intensity. Prior work often relies on weather simulation or generic augmentation and treats geometry and reflectance as a unified signal, overlooking their distinct degradation mechanisms. We propose an explicit geometry–reflectance domain shift modeling framework for robust LiDAR segmentation under both domain generalization (DG) and unsupervised domain adaptation (UDA). At the <Emphasis Type="BoldItalic">model level</Emphasis>, we introduce the Geometry–Reflectance Collaboration Network (GRCNet), which separately encodes geometry and reflectance and fuses them via robust multi-level feature collaboration to suppress weather-induced corruption while preserving complementary semantics. When unlabeled target data is available, we extend the same factorization to the <Emphasis Type="BoldItalic">data level</Emphasis> with a novel Target-Driven Domain-Augmented (TDDA) strategy, which explicitly models and transfers target-domain shifts along reflectance statistics and geometric degradation, without complex physical simulation. Extensive experiments on multiple benchmarks show consistent improvements over state-of-the-art methods in both DG and UDA settings. Code will be available at <a href="https://github.com/momoyo126/GRCNet.git">https://github.com/momoyo126/GRCNet.git</a>.</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

Explicit Geometry-Reflectance Domain Shift Modeling for Robust LiDAR Segmentation in Adverse Weather

  • 期刊论文

  • 发表日期：

    2026年7月24日

查看全文
  • Longyu Yang,
  • Shangbo Yuan,
  • Lu Zhang,
  • Jun Liu,
  • Heng tao Shen,
  • Xiaofeng Zhu,
  • Ping Hu

摘要

LiDAR semantic segmentation degrades severely in adverse weather due to heterogeneous, modality-dependent domain shifts in geometric structure and reflectance intensity. Prior work often relies on weather simulation or generic augmentation and treats geometry and reflectance as a unified signal, overlooking their distinct degradation mechanisms. We propose an explicit geometry–reflectance domain shift modeling framework for robust LiDAR segmentation under both domain generalization (DG) and unsupervised domain adaptation (UDA). At the model level, we introduce the Geometry–Reflectance Collaboration Network (GRCNet), which separately encodes geometry and reflectance and fuses them via robust multi-level feature collaboration to suppress weather-induced corruption while preserving complementary semantics. When unlabeled target data is available, we extend the same factorization to the data level with a novel Target-Driven Domain-Augmented (TDDA) strategy, which explicitly models and transfers target-domain shifts along reflectance statistics and geometric degradation, without complex physical simulation. Extensive experiments on multiple benchmarks show consistent improvements over state-of-the-art methods in both DG and UDA settings. Code will be available at https://github.com/momoyo126/GRCNet.git.