面向微小密集目标分割的结构感知表征蒸馏
Structure-Aware Representation Distillation for Tiny-Dense Object Segmentation
把教师特征的结构重要性（边界显著性、几何复杂度、局部特征变化）做成加权图来引导蒸馏损失，可直接迁移到任何密集预测任务的师生对齐中，替代逐像素 mask 模仿。
原文依据 transfers structural knowledge from a large teacher to a compact student via feature-space alignment rather than mask imitation
摘要译文
由模型翻译 · 术语统一 · 请对照原文核对关键结论包含大量微小目标的密集场景对分割模型构成根本性挑战，微小的定位误差就可能显著降低下游测量精度。本文提出结构感知表征蒸馏（SARD），这是一个与教师模型兼容的框架，通过特征空间对齐而非掩码模仿，将大教师模型的结构知识迁移到紧凑的学生模型。SARD 通过融合边界显著性、几何复杂度和局部特征变化构建结构重要性图，并用其引导一个统一表征损失，该损失结合了特征一致性与分布对齐。这促使学生模型将容量分配到几何信息丰富的区域，同时保留全局上下文。在 Cityscapes、ADE20K 以及具有挑战性的岩石破碎基准（RockFrag）上的实验表明，SARD 在强蒸馏基线上持续提升平均交并比和边界交并比；在 RockFrag 上，SARD 使 Swin-T 学生模型相比 CWD 提升 +4.3 平均交并比和 +6.7 边界交并比。由 Swin-L 教师模型蒸馏得到的 ResNet-50 学生模型实现了最高 7.7 倍的参数量缩减和 9 倍的吞吐量提升，且除学生网络外无额外推理开销，表明结构感知表征蒸馏对微小密集分割既有效又高效。代码见：https://github.com/liuuuuuuxuesong/SARD。
英文原文摘要 1434 字符
Dense scenes containing numerous tiny objects pose a fundamental challenge for segmentation models, where small localization errors can significantly degrade downstream measurements. We present Structure-Aware Representation Distillation (SARD), a teacher-compatible framework that transfers structural knowledge from a large teacher to a compact student via feature-space alignment rather than mask imitation. SARD builds a structure-importance map by integrating boundary salience, geometric complexity, and local feature variation, and uses it to guide a unified representation loss that combines feature consistency with distribution alignment. This encourages the student to allocate capacity to geometrically informative regions while preserving global context. Experiments on Cityscapes, ADE20K, and a challenging rock fragmentation benchmark (RockFrag) show that SARD consistently improves both mIoU and boundary IoU over strong distillation baselines; on RockFrag, SARD improves a Swin-T student over CWD by +4.3 mIoU and +6.7 bIoU. A ResNet-50 student distilled from a Swin-L teacher achieves up to 7.7 times parameter reduction and 9 times higher throughput than the teacher, with no additional inference overhead beyond the student network, demonstrating that structure-aware representation distillation is effective and efficient for tiny-dense segmentation. Code is available at: https://github.com/liuuuuuuxuesong/SARD.