MetaSeg: Content-Aware Meta-Net for Omni-Supervised Semantic Segmentation.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Noisy labels in pseudo-segmentation data hinder semantic segmentation model optimization.
• Existing methods often use complex handcrafted losses and hyperparameter tuning, leading to poor generalization.
• A passive approach to noise tolerance is less effective than actively identifying and ignoring noisy regions.

Purpose of the Study:

• To develop a novel meta-learning-based semantic segmentation method (MetaSeg) to actively identify and suppress noisy regions.
• To improve the accuracy and efficiency of semantic segmentation models trained with imperfect labels.
• To provide a more feasible solution for omni-supervised semantic segmentation.

Main Methods:

• Introduced MetaSeg, a semantic segmentation method utilizing a content-aware meta-net (CAM-Net) as a noise indicator.
• CAM-Net generates pixel-wise weights to suppress noisy regions and highlight clean regions using hybrid features.
• Implemented a decoupled training strategy to optimize different layers of large segmentation models efficiently.

Main Results:

• MetaSeg effectively suppresses noisy regions and highlights clean regions, guiding segmentation model optimization.
• The decoupled training strategy addresses the time-consuming training issue in meta-learning for large models.
• Achieved superior performance across object, medical, remote sensing, and human segmentation tasks, nearing fully supervised settings.

Conclusions:

• MetaSeg offers a promising new direction for omni-supervised semantic segmentation by actively managing noisy labels.
• The method demonstrates robust performance and improved generalization compared to traditional noise-handling techniques.
• This work paves the way for more efficient and accurate semantic segmentation using weakly annotated data.