AI 용어집
인공지능 완전 사전
Query-based Detection
Detection paradigm where learned queries (embeddings) interact with image features through an attention mechanism to directly predict bounding boxes and object classes.
Object Queries
Positional learning vectors in DETR architectures that act as 'slots' for each potential object to detect, guiding the model toward specific predictions.
Bipartite Matching Loss
Loss function used in DETR that finds the optimal one-to-one matching between predictions and ground truths using the Hungarian algorithm, ensuring unique assignment for each object.
Multi-Scale Feature Pyramid
Structure in transformer detectors that combines features from different resolutions to improve detection of objects of varying sizes, often through cross-scale attention mechanisms.
Anchor-Free Detection
Detection approach that eliminates the use of predefined anchor boxes, a key feature of transformer architectures that directly predict bounding boxes.
Set Prediction
Formulation of object detection as an unordered set prediction problem, where the model simultaneously predicts all objects without predefined order.
Class-Agnostic Detection
Approach where object localization and classification are decoupled, often used in transformer detectors to improve generalization.
Vision Transformer (ViT) Backbone
Use of pre-trained ViTs as feature extractors for transformer detectors, offering powerful and contextual image representation.
DINO (DETR with Improved deNoising anchOr)
Advanced detection architecture that combines denoised queries and anchors to improve the performance and convergence speed of transformer detectors.
Denoising Training
Training strategy where the model learns to reconstruct ground truths from noised versions, improving the robustness and convergence of transformer detectors.
Query-to-Instance Attention
Specialized attention mechanism where each object query focuses on the relevant features of a specific instance in the image.
One-to-Many Label Assignment
Alternative assignment strategy in some transformer detectors where a ground truth can be assigned to multiple predictions to improve training.