KI-Glossar
Das vollständige Wörterbuch der Künstlichen Intelligenz
Segmenter
Semantic segmentation model based on a pure transformer architecture, designed to efficiently capture long-range contextual relationships between pixels.
Learnable Token
Randomly initialized embedding vector learned during training, used in transformer decoders to aggregate contextual information and predict segmentation classes.
Segmentation Transformer Decoder
Module that reconstructs a high-resolution segmentation map from encoder features, using attention mechanisms to refine predictions pixel by pixel.
SegFormer
Efficient and simple segmentation architecture based on a hierarchical transformer encoder and lightweight decoder (All-MLP), designed for better performance with fewer parameters.
Masked Autoencoding (MAE)
Self-supervised pre-training strategy where large portions of an image are masked and the model learns to reconstruct them, improving contextual understanding for segmentation.
Query-Based Segmentation
Paradigm where a fixed set of learnable query vectors is used to query image features and directly generate segmentation masks.
Hierarchical Windowing
Technique in vision transformers that divides the image into windows at different scales and hierarchically merges them to capture both local details and global context.
Class Embedding
Learned vector representation for each semantic category, used in transformer decoders to guide pixel classification and improve prediction consistency.