Weakly Supervised Semantic Segmentation

Weakly supervised semantic segmentation (WSSS) aims to train accurate image segmentation models using only image-level labels, significantly reducing the need for expensive pixel-level annotations. Current research focuses on improving the quality of pseudo-labels generated from these weak labels, often employing vision transformers (ViTs) and convolutional neural networks (CNNs) in conjunction with techniques like contrastive learning, domain adaptation, and multi-modal approaches (e.g., incorporating text embeddings). Advances in WSSS are crucial for expanding the applicability of semantic segmentation to diverse domains where large, fully annotated datasets are unavailable, impacting fields such as medical image analysis and autonomous driving.