Scholarlink

<p>The objective of referring 3D instance segmentation (3DRIS) is to precisely identify and isolate the specific instance (i.e., referent) in a given set of point clouds referred to by a given expression. Existing techniques generate a set of instance proposals and then rely on similarity measures to locate the referent. However, they usually directly calculate the similarity between the expression and instance proposals, ignoring critical aspects such as object perception and relation reasoning. To address these limitations, we propose the Chain of Perception (CoP) framework, which progressively enhances perception and reasoning capabilities. The CoP framework consists of three primary components. Firstly, we introduce an object perception module that identifies all objects mentioned in the referring expression. Secondly, we propose a relation reasoning module that uses spatial information and relationship words to model relationships between instances. Finally, we present a cross-modal interaction module that facilitates mutual interaction between instances and expressions. Additionally, we also introduce a novel task called <i>X-3DRIS</i>, which involves segmenting the referent without access to the target object names. Our experiments conducted on the ScanRefer and X-ScanRefer benchmarks demonstrate the superiority of our proposed approach over existing methods, achieving state-of-the-art performance on both tasks.</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

CoP: Chain of Perception for Referring 3D Instance Segmentation

  • 期刊论文

  • 发表日期：

    2026年3月7日

查看全文
  • Yiwei Ma,
  • Jiayi Ji,
  • Zhipeng Qian,
  • Xiaoshuai Sun,
  • Rongrong Ji

摘要

The objective of referring 3D instance segmentation (3DRIS) is to precisely identify and isolate the specific instance (i.e., referent) in a given set of point clouds referred to by a given expression. Existing techniques generate a set of instance proposals and then rely on similarity measures to locate the referent. However, they usually directly calculate the similarity between the expression and instance proposals, ignoring critical aspects such as object perception and relation reasoning. To address these limitations, we propose the Chain of Perception (CoP) framework, which progressively enhances perception and reasoning capabilities. The CoP framework consists of three primary components. Firstly, we introduce an object perception module that identifies all objects mentioned in the referring expression. Secondly, we propose a relation reasoning module that uses spatial information and relationship words to model relationships between instances. Finally, we present a cross-modal interaction module that facilitates mutual interaction between instances and expressions. Additionally, we also introduce a novel task called X-3DRIS, which involves segmenting the referent without access to the target object names. Our experiments conducted on the ScanRefer and X-ScanRefer benchmarks demonstrate the superiority of our proposed approach over existing methods, achieving state-of-the-art performance on both tasks.