Scholarlink

<p>Amodal instance segmentation seeks to reconstruct the entire shape of occluded objects, including the invisible parts. It can provide substantial advantages to subsequent tasks like semantic mapping and object tracking. Although existing methods have generated encouraging results, better performance can be achieved by addressing the current challenges of underutilized correlations between object parts and unresolved conflicts between prediction masks. In this work, we introduce MAS-Net, a powerful two-stage network for amodal instance segmentation, to predict category-agnostic occluder, visible, invisible, and amodal masks of occluded objects. Two vital schemes are introduced in MAS-Net to remedy the current shortcomings. We propose a multi-level modeling (MLM) scheme to analyze object occlusion by defining the stages of feature integration and the sequence of predictions. Following this scheme, we construct a corresponding MLM head to enable MAS-Net to accomplish category-agnostic mask prediction progressively and efficiently. In addition, we propose a mask merging scheme to alleviate the conflicts between prediction masks during inference. We empirically validate MAS-Net on four amodal datasets, i.e., COCOA-cls, KINS, D2SA and CWALT. Experimental results demonstrate the excellence of MAS-Net by achieving superior performance compared to most current state-of-the-art methods. The code is public at <a href="https://github.com/jiaoZ7688/MAS-Net">https://github.com/jiaoZ7688/MAS-Net</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

Multi-level modeling-based amodal instance segmentation network

  • 期刊论文

  • 发表日期：

    2025年9月16日

查看全文
  • Bohan Yang,
  • Jiao Zhan,
  • Jingnan Liu,
  • Chi Guo

摘要

Amodal instance segmentation seeks to reconstruct the entire shape of occluded objects, including the invisible parts. It can provide substantial advantages to subsequent tasks like semantic mapping and object tracking. Although existing methods have generated encouraging results, better performance can be achieved by addressing the current challenges of underutilized correlations between object parts and unresolved conflicts between prediction masks. In this work, we introduce MAS-Net, a powerful two-stage network for amodal instance segmentation, to predict category-agnostic occluder, visible, invisible, and amodal masks of occluded objects. Two vital schemes are introduced in MAS-Net to remedy the current shortcomings. We propose a multi-level modeling (MLM) scheme to analyze object occlusion by defining the stages of feature integration and the sequence of predictions. Following this scheme, we construct a corresponding MLM head to enable MAS-Net to accomplish category-agnostic mask prediction progressively and efficiently. In addition, we propose a mask merging scheme to alleviate the conflicts between prediction masks during inference. We empirically validate MAS-Net on four amodal datasets, i.e., COCOA-cls, KINS, D2SA and CWALT. Experimental results demonstrate the excellence of MAS-Net by achieving superior performance compared to most current state-of-the-art methods. The code is public at https://github.com/jiaoZ7688/MAS-Net.