Table 2 An overview of fully-supervised and semi-supervised segmentation methods for comparison.

FCN⁶⁶: utilizes fully convolutional layers to realize dense prediction for arbitrary-sized images.

DCT⁴⁵: employs one network to ensure consistency across different views of a given sample.

PSPNet⁶⁷: Utilizes global context aggregation through pyramid pooling for complicated scene parsing.

CPS⁴³: enforces consistency between two segmentation networks initialized differently.

DeepLabV3+³⁶: introduced the atrous spatial convolutional pyramid (ASPP) to enhance the multi-scale contextual information.

UAMT⁴⁶: encourages consistent predictions under different perturbations and estimates uncertainty to learn from unlabeled data.

DANet⁶⁸: enhances segmentation by adaptively integrating semantic dependencies in spatial and channel dimensions via the self-attention mechanism.

UCC⁴⁴: employs a shared encoder with dual decoders and enforces consistency between the decoders with data augmentations.

OCRNet⁶⁹: introduces object-contextual representations for semantic segmentation, leveraging pixel-object relationships to augment pixel representations.

UAPS⁵: dynamically blends pseudo-labels from multi-head outputs during a single forward pass for uncertainty regularization.

SegFormer⁴⁰: presents a streamlined semantic segmentation framework by integrating Transformers with lightweight MLP decoders.