跳转至

文章背景与核心概要

当前的开放词汇实例分割(OVIS)和开放集全景分割(OSPS)方法在面对带噪的伪掩码、有限的视觉-文本基础对齐以及处理同义词或词表外(OOV)词汇时,往往表现出明显的局限性。本文针对这些痛点,提出了一种新颖的多模态框架,能够在无需人工标注的情况下自动生成高质量的监督信号。

通过整合 Grounded SAMLLaVACLIP,该框架构建了鲁棒的伪分割掩码、描述性标题以及语义对齐的同义词集。模型通过三种专门设计的训练目标进一步优化——包括扩展基础损失(extended grounding loss)、语义一致性损失(semantic consistency loss)和生成式标题重建(generative caption reconstruction),从而在 COCO 数据集上实现了最先进的性能表现。


从多模态伪标签中学习:实现鲁棒的开放词汇实例与全景分割

作者: Duy Tran Thanh, Yeejin Lee, Byeongkeun Kang
发布时间: 2026年8月12日
期刊: Neurocomputing (2026)
arXiv: 2608.11681


摘要 (Summary)

本文针对当前开放词汇实例分割(OVIS)和开放集全景分割(OSPS)方法的局限性进行了研究,这些方法通常难以应对带噪的伪掩码、有限的视觉-文本基础对齐,以及处理同义词或词表外(OOV)词汇的困难。

This paper addresses the limitations of current Open-Vocabulary Instance Segmentation (OVIS) and Open-Set Panoptic Segmentation (OSPS) methods, which often struggle with noisy pseudo-masks, limited visual-textual grounding, and difficulties in handling synonyms or out-of-vocabulary (OOV) terms.

作者提出了一种新颖的多模态框架,在无需人工标注的情况下自动生成高质量的监督信号。通过集成 Grounded SAMLLaVACLIP,该框架构建了鲁棒的伪分割掩码、描述性标题以及语义对齐的同义词集。模型通过三个专门的训练目标进一步优化——扩展基础损失、语义一致性损失和生成式标题重建,从而在 COCO 数据集上取得了最先进的性能。

The authors propose a novel multimodal framework that automates the generation of high-quality supervision without manual annotation. By integrating Grounded SAM, LLaVA, and CLIP, the framework constructs robust pseudo-segmentation masks, descriptive captions, and semantically aligned synonym sets. The model is further refined through three specialized training objectives—extended grounding loss, semantic consistency loss, and generative caption reconstruction—resulting in state-of-the-art performance on the COCO dataset.


核心贡献 (Key Contributions)

  • 自动多模态监督: 利用预训练的视听语言模型生成伪标签,有效地避开了对穷举式人工标注的需求。

    • Automated Multimodal Supervision: Leverages pre-trained vision-language models to generate pseudo-labels, effectively bypassing the need for exhaustive human annotation.
  • 增强的语义对齐: 引入了 CLIP引导的同义词过滤基于GPT的标题重建,以弥合视觉特征与文本描述之间的鸿沟。

    • Enhanced Semantic Alignment: Introduces CLIP-guided synonym filtering and GPT-based caption reconstruction to bridge the gap between visual features and textual descriptions.
  • 鲁棒的训练目标:

    • 扩展基础损失: 结合了视觉基础同义词以提高物体识别能力。
    • 语义一致性损失: 确保视觉片段与其对应文本之间的连贯映射。
    • 生成式标题重建损失: 提升模型描述和识别未见类别的能力。
      • Robust Training Objectives:
        • Extended Grounding Loss: Incorporates visually grounded synonyms to improve object recognition.
        • Semantic Consistency Loss: Ensures coherent mapping between visual segments and their textual counterparts.
        • Generative Caption Reconstruction Loss: Improves the model's ability to describe and identify unseen categories.
  • 基准性能: 在 COCO 数据集的 OVIS 和 OSPS 基准测试中,展现出对现有最先进方法的显著提升。

    • Benchmark Performance: Demonstrated significant improvements over existing state-of-the-art methods on both OVIS and OSPS benchmarks using the COCO dataset.

元数据 (Metadata)

字段 (Field) 详情 (Details)
主要学科 (Primary Subject) 计算机视觉与模式识别 (cs.CV)
次要学科 (Secondary Subjects) 人工智能 (cs.AI), 多媒体 (cs.MM)
全文 (Full-Text) 查看 PDF (View PDF)
源代码 (Source Code) TeX 源码 (TeX Source)
DOI 10.48550/arXiv.2608.11681