Scholarlink

<p>Robust object detection in real-world scenes remains challenging when structural occlusion, adverse environmental appearance shifts, and fine-grained inter-class ambiguity occur simultaneously. In inland waterway monitoring, these degradation factors are often strongly coupled and mutually reinforcing: vessel occlusion leads to structural information loss, environmental conditions such as fog, rain, and low illumination further weaken the limited visible features, while the high visual similarity among cargo vessel categories amplifies recognition ambiguity. However, existing methods typically address individual degradation factors in isolation, making them insufficient for handling their combined effects. To address this issue, we propose VCR-ShipFormer, a visibility-conditioned Transformer framework for robust vessel detection under compound visual degradation. Built upon Deformable DETR, the proposed framework jointly models environment-invariant representation learning, visibility-aware part reasoning, and fine-grained contrastive discrimination to improve detection robustness in challenging inland waterway environments. In addition, we introduce SeaShips-OccEnv, the first inland vessel detection benchmark jointly annotated with occlusion severity, part visibility, and multi-weather conditions, comprising 7000 images spanning six vessel categories and five environmental domains. Extensive experiments demonstrate that VCR-ShipFormer achieves 88.1% mAP at 20 FPS, outperforming RT-DETR-X by 8.43 and 8.11 percentage points under foggy and nighttime conditions, respectively, while exhibiting a cross-environment standard deviation of only 1.00%, indicating superior robustness and generalization. Code, trained models, dataset annotations, and evaluation protocols are publicly available at <a href="https://github.com/lyuzhetao/lyuzhetao.git">https://github.com/lyuzhetao/lyuzhetao.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

VCR-ShipFormer: visibility-conditioned transformer learning for robust vessel detection in degraded inland waterway scenes

  • 期刊论文

  • 发表日期：

    2026年7月8日

查看全文
  • Zhetao Lyu,
  • Chenbin Xu,
  • Jianling Cai,
  • Chaoling Li,
  • Yutao Jiang

摘要

Robust object detection in real-world scenes remains challenging when structural occlusion, adverse environmental appearance shifts, and fine-grained inter-class ambiguity occur simultaneously. In inland waterway monitoring, these degradation factors are often strongly coupled and mutually reinforcing: vessel occlusion leads to structural information loss, environmental conditions such as fog, rain, and low illumination further weaken the limited visible features, while the high visual similarity among cargo vessel categories amplifies recognition ambiguity. However, existing methods typically address individual degradation factors in isolation, making them insufficient for handling their combined effects. To address this issue, we propose VCR-ShipFormer, a visibility-conditioned Transformer framework for robust vessel detection under compound visual degradation. Built upon Deformable DETR, the proposed framework jointly models environment-invariant representation learning, visibility-aware part reasoning, and fine-grained contrastive discrimination to improve detection robustness in challenging inland waterway environments. In addition, we introduce SeaShips-OccEnv, the first inland vessel detection benchmark jointly annotated with occlusion severity, part visibility, and multi-weather conditions, comprising 7000 images spanning six vessel categories and five environmental domains. Extensive experiments demonstrate that VCR-ShipFormer achieves 88.1% mAP at 20 FPS, outperforming RT-DETR-X by 8.43 and 8.11 percentage points under foggy and nighttime conditions, respectively, while exhibiting a cross-environment standard deviation of only 1.00%, indicating superior robustness and generalization. Code, trained models, dataset annotations, and evaluation protocols are publicly available at https://github.com/lyuzhetao/lyuzhetao.git.