Scholarlink

<p>Instance segmentation, providing pixel-level object delineation with instance discrimination, has become a cornerstone task in computer vision with applications spanning autonomous driving, medical diagnosis, and robotic manipulation. Despite remarkable progress driven by deep learning, significant challenges persist in handling occlusion, multi-scale variation, boundary precision, and computational efficiency. This comprehensive review systematically categorizes and analyzes instance segmentation algorithms across three evolutionary paradigms: CNN-based methods (two-stage and single-stage), transformer-based architectures, and foundation models. Through quantitative meta-analysis of 100+ papers spanning 2014–2024, we establish performance-efficiency trade-off boundaries and identify critical failure modes across standard benchmarks (MS COCO, Cityscapes, and LVIS). Our key contributions include: (1) a novel taxonomy integrating architectural design, learning paradigms, and application scenarios; (2) comprehensive performance analysis revealing an approximately 15–18% average precision gap between state-of-the-art methods and human-level segmentation; (3) quantitative evaluation of emerging techniques including prompt-based segmentation, weakly supervised learning, and domain adaptation; and (4) identification of five critical research frontiers with concrete technical roadmaps. We conclude that future breakthroughs will emerge from foundation model scaling, efficient architecture search, and hybrid supervision strategies, which may contribute to narrowing the remaining performance gap, though the precise timeline depends on the pace of paradigm-level innovations.</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

Deep Learning-Based Instance Segmentation: A Comprehensive Review of Algorithms, Challenges, and Future Directions

  • 期刊论文

  • 发表日期：

    2026年4月1日

查看全文
  • Jiacheng Lou,
  • Sergei Shavetov,
  • Xuecheng Wen,
  • Zhidong Li,
  • Xiang Zhang,
  • Chunhong Yuan

摘要

Instance segmentation, providing pixel-level object delineation with instance discrimination, has become a cornerstone task in computer vision with applications spanning autonomous driving, medical diagnosis, and robotic manipulation. Despite remarkable progress driven by deep learning, significant challenges persist in handling occlusion, multi-scale variation, boundary precision, and computational efficiency. This comprehensive review systematically categorizes and analyzes instance segmentation algorithms across three evolutionary paradigms: CNN-based methods (two-stage and single-stage), transformer-based architectures, and foundation models. Through quantitative meta-analysis of 100+ papers spanning 2014–2024, we establish performance-efficiency trade-off boundaries and identify critical failure modes across standard benchmarks (MS COCO, Cityscapes, and LVIS). Our key contributions include: (1) a novel taxonomy integrating architectural design, learning paradigms, and application scenarios; (2) comprehensive performance analysis revealing an approximately 15–18% average precision gap between state-of-the-art methods and human-level segmentation; (3) quantitative evaluation of emerging techniques including prompt-based segmentation, weakly supervised learning, and domain adaptation; and (4) identification of five critical research frontiers with concrete technical roadmaps. We conclude that future breakthroughs will emerge from foundation model scaling, efficient architecture search, and hybrid supervision strategies, which may contribute to narrowing the remaining performance gap, though the precise timeline depends on the pace of paradigm-level innovations.