Scholarlink

<p>Scene graph generation (SGG) aims to identify object relationships for structured representations. However, long-tailed distributions often cause biased coarse-grained recognition. We propose a plug-and-play Cross-region node Generation and Relationship Re-Filtration (CGRRF) framework. Specifically, the CRNG module identifies interaction regions to generate cross sub-nodes, introducing a local fine-grained branch complementary to global representations. Crucially, this branch is supervised by the RRF module, which leverages dataset statistics to guide the mined local information to specifically focus on discriminating tail predicates for effective debiasing. Finally, local and global features are fused for precise scene graph construction. Experiments on Visual Genome show CGRRF consistently improves mR@K across baselines. Moreover, since our framework introduces additional region-level interaction nodes over large object-pair combinations, it naturally benefits from parallel and HPC-enabled acceleration, making it well aligned with high-throughput visual reasoning applications. The code is publicly available at <a href="https://github.com/a742136653/CGRRF">https://github.com/a742136653/CGRRF</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

Capturing local information from cross-region for unbiased scene graph generation

  • 期刊论文

  • 发表日期：

    2026年3月6日

查看全文
  • Yongfeng Dong,
  • Kunyu Li,
  • Hao Cheng,
  • Dong Han,
  • Linhao Li

摘要

Scene graph generation (SGG) aims to identify object relationships for structured representations. However, long-tailed distributions often cause biased coarse-grained recognition. We propose a plug-and-play Cross-region node Generation and Relationship Re-Filtration (CGRRF) framework. Specifically, the CRNG module identifies interaction regions to generate cross sub-nodes, introducing a local fine-grained branch complementary to global representations. Crucially, this branch is supervised by the RRF module, which leverages dataset statistics to guide the mined local information to specifically focus on discriminating tail predicates for effective debiasing. Finally, local and global features are fused for precise scene graph construction. Experiments on Visual Genome show CGRRF consistently improves mR@K across baselines. Moreover, since our framework introduces additional region-level interaction nodes over large object-pair combinations, it naturally benefits from parallel and HPC-enabled acceleration, making it well aligned with high-throughput visual reasoning applications. The code is publicly available at https://github.com/a742136653/CGRRF.