IGL-DT: Iterative Global-Local Feature Learning with Dual-Teacher Semantic Segmentation Framework under Limited Annotation Scheme

Event: CVPR Nashville 2025 · Duration: 8 min · ▶ Watch on YouTube

Abstract

Semi-supervised semantic segmentation aims to improve model performance using a small set of labeled images and a large pool of unlabeled data. Current methods often rely on a single model backbone or generate noisy pseudo-labels, struggling to effectively learn both high-level context and local details. This paper proposes IGL-DT, an iterative global-local feature learning framework with a dual-teacher strategy. IGL-DT addresses these issues by combining a Transformer and a CNN to extract complementary global and local knowledge, guided by Global Context Learning, Local Regional Learning, and Discrepancy Learning.

Speakers

  • Dinh Dai Quan Tran — National Chung Cheng University, AI VIET NAM
  • Hoang-Thien Nguyen — National Chung Cheng University, AI VIET NAM
  • Thanh-Huy Nguyen — National Chung Cheng University, AI VIET NAM
  • Gia-Van To — National Chung Cheng University, AI VIET NAM
  • Tien-Huy Nguyen — National Chung Cheng University, AI VIET NAM
  • Quan Nguyen — National Chung Cheng University, AI VIET NAM

Talks (1)

  • 00:00:00 — Dinh Dai Quan Tran: IGL-DT: Iterative Global-Local Feature Learning with Dual-Teacher Semantic Segmentation Framework under Limited Annotation Scheme
    • Presents IGL-DT, a semi-supervised semantic segmentation framework that iteratively learns global and local features using a dual-teacher strategy to improve performance under limited annotation.

Key Takeaways

  • IGL-DT is a novel semi-supervised semantic segmentation framework that leverages a dual-teacher strategy with complementary Transformer (global) and CNN (local) backbones.
  • The framework incorporates Global Context Learning, Local Regional Learning, and Discrepancy Learning to effectively guide the student model and prevent overfitting.
  • IGL-DT achieves state-of-the-art quantitative results on Pascal VOC and Cityscapes datasets, outperforming strong baselines.
  • Visual comparisons demonstrate improved segmentation accuracy, especially for challenging cases like small objects, boundaries, and occlusions.
  • Ablation studies confirm the unique contribution of each loss component and the synergistic benefits of combining global and local feature learning with discrepancy learning.

Methods / Models / Datasets Mentioned

  • IGL-DT
  • SwinUnet
  • ResNet
  • CPS
  • PS-MT
  • U2PL
  • ST++
  • CCVC
  • CorrMatch
  • Dual-Teacher
  • Pascal VOC
  • Cityscapes

Topics

Semi-supervised semantic segmentation · Dual-teacher framework · Global context learning · Local regional learning · Discrepancy learning · Transformer · CNN · Limited annotation · Pseudo-labeling · Iterative learning

Notes

Open for commentary — connections to other work, critiques, follow-up reading.