Semantic Segmentation

Semantic segmentation predicts a category for each pixel, supporting applications such as autonomous driving, medical imaging, and remote sensing. Common approaches include CNN and vision-transformer encoders with multi-scale decoders, while foundation models, vision-language prompts, and diffusion features are increasingly adapted for dense prediction. To reduce annotation demands, researchers use weak or sparse labels, pseudo-labeling, synthetic data, few-shot learning, and domain adaptation; RGB-D, thermal, radar, and other sensors provide complementary cues. Key challenges are reliable transfer across domains and conditions, accurate boundaries and rare classes, and balancing accuracy with calibration and inference cost.