KI-Glossar
Das vollständige Wörterbuch der Künstlichen Intelligenz
Bipartite Matching
One-to-one matching technique between predictions and ground truths using the Hungarian algorithm to optimize assignment during training.
Set Prediction Loss
Loss function combining classification and bounding box regression with bipartite matching to directly optimize the set of predictions.
NMS-Free Detection
Approach that eliminates the Non-Maximum Suppression step by directly producing a unique set of predictions without redundancy.
Query-to-Scene Attention
Process where each object query selectively attends to relevant regions of the image to refine its prediction.
Multi-Scale Features
Feature extraction at different spatial resolutions to improve detection of objects of varying sizes in DETR.
Class Prediction Head
Final module that transforms query embeddings into probability scores for each possible object class.
Auxiliary Loss
Additional loss added to intermediate decoder layers to stabilize training and accelerate convergence.