Skip to main content
QUICK REVIEW

[论文解读] MATANet: A Multi-context Attention and Taxonomy-Aware Network for Fine-Grained Underwater Recognition of Marine Species

Donghwan Lee, Byeongjin Kim|arXiv (Cornell University)|Jan 7, 2026
Advanced Neural Network ApplicationsComputer Science被引用 0
一句话总结

MATANet 将多上下文环境注意力与分类学感知学习相结合,在水下与遥感数据集上实现细粒度海洋生物识别的SOTA。

ABSTRACT

Fine-grained classification of marine animals supports ecology, biodiversity and habitat conservation, and evidence-based policy-making. However, existing methods often overlook contextual interactions from the surrounding environment and insufficiently incorporate the hierarchical structure of marine biological taxonomy. To address these challenges, we propose MATANet (Multi-context Attention and Taxonomy-Aware Network), a novel model designed for fine-grained marine species classification. MATANet mimics expert strategies by using taxonomy and environmental context to interpret ambiguous features of underwater animals. It consists of two key components: a Multi-Context Environmental Attention Module (MCEAM), which learns relationships between regions of interest (ROIs) and their surrounding environments, and a Hierarchical Separation-Induced Learning Module (HSLM), which encodes taxonomic hierarchy into the feature space. MATANet combines instance and environmental features with taxonomic structure to enhance fine-grained classification. Experiments on the FathomNet2025, FAIR1M, and LifeCLEF2015-Fish datasets demonstrate state-of-the-art performance. The source code is available at: https://github.com/dhlee-work/fathomnet-cvpr2025-ssl

研究动机与目标

  • 推动利用环境上下文和分类学知识提升细粒度海洋生物分类的识别准确性。
  • 引入一个将上下文注意力与分层学习相结合的双组件架构。
  • 在 incorporating habitat context 的同时实现与分类学层级的一致语义。
  • 在多个数据集上展示最先进的性能,包括 FathomNet2025、FAIR1M 和 LifeCLEF2015-Fish。

提出的方法

  • 提出 MATANet,包含两大模块:(i) Multi-Context Environmental Attention Module (MCEAM),在 ROI 嵌入与多尺度上下文区域嵌入之间执行跨注意力;(ii) Hierarchical Separation-Induced Learning Module (HSLM),通过逐层辅助分类器将分类层级嵌入特征空间。
  • 使用 Vision Transformer (ViT) 主干编码 ROI 与上下文区域;通过拼接和投影层进行融合,得到统一嵌入 z。
  • 对 ROI 作为查询,上下文区域作为键/值,执行跨注意力以捕捉 ROI-上下文交互。
  • 通过逐级辅助分类器施加分层监督,计算交叉熵损失,并与最终分类损失组合。
  • 在训练中使用 FathomNet API 的真实标签,并在缺失低级标签时插值到更高级别的分类。
  • 端到端训练,损失函数为 L_total = L_cls + L_hier,平衡对象级准确性与分类学一致性。
Figure 1: Different visual contexts for marine species recognition. The ROI image shows a sea anemone (family: Hormathiidae). Its taxonomic identity becomes more evident when the surrounding organisms and habitat features visible in the context and full-context images are taken into account.
Figure 1: Different visual contexts for marine species recognition. The ROI image shows a sea anemone (family: Hormathiidae). Its taxonomic identity becomes more evident when the surrounding organisms and habitat features visible in the context and full-context images are taken into account.

实验结果

研究问题

  • RQ1引入多尺度生态上下文是否能超越以 ROI 为中心的方法,提升细粒度海洋生物识别准确性?
  • RQ2通过逐级辅助分类器编码分层分类结构是否能增强语义一致性与分类准确性?
  • RQ3上下文尺度和分层监督对嵌入空间结构及在不同海洋数据集上的最终性能有何影响?
  • RQ4MATANet 在具有细粒度分类标签的水下与遥感数据集上的泛化能力如何?

主要发现

  • MATANet 在 FathomNet2025、FAIR1M 和 LifeCLEF2015-Fish 数据集上实现了最先进的性能。
  • 消融实验表明多尺度上下文(3×、5×、全尺度)相较单一尺度上下文提升了性能。
  • 引入 Hierarchical Separation-Induced Learning (HSLM) 显著提升了分层一致性并降低了分层距离(HD)。
  • 更大尺寸的 ViT 主干(如 ViT-L)和 MCEAM-HSLM 配置获得最佳结果,相较于仅 ROI 的基线有显著提升。
  • 可视化分析(t-SNE)显示 HSLM 产生的嵌入在分类和分类学上更具植物性/分类学的一致性。
  • 跨注意力可视化展示了有意义的 ROI-上下文交互,聚焦于与栖息地相关的特征。
Figure 2: Overview of MATANet. The model processes ROI and multi-scale contextual images using ViT embeddings. The Multi-Context Environmental Attention Module (MCEAM) applies cross-attention between the ROI and contextual regions to integrate their features. The Hierarchical Separation-Induced Lear
Figure 2: Overview of MATANet. The model processes ROI and multi-scale contextual images using ViT embeddings. The Multi-Context Environmental Attention Module (MCEAM) applies cross-attention between the ROI and contextual regions to integrate their features. The Hierarchical Separation-Induced Lear

阅读本论文

在 Nubint 阅读器中打开原文,边读边就任意段落提问。

在 Nubint 中阅读本论文

本解读由 AI 根据论文原文生成,引用前请核对原文。