AI-ordlista
Den kompletta ordlistan över AI
Cascade regression
Iterative process where each detector in the cascade refines the predictions of the previous detector, using progressively higher IoU thresholds to improve accuracy.
R-CNN (Region-based CNN)
Family of object detection architectures based on regions of interest extracted from images, combining region proposals with convolutional neural networks for classification and localization.
Adaptive IoU thresholds
Set of increasing IoU thresholds (typically 0.5, 0.6, 0.7) used in Cascade R-CNN to train specialized detectors with different levels of detection quality.
Detector head
Classification and regression module placed after the backbone, responsible for predicting object classes and refining bounding box coordinates for each ROI.
Multi-stage training
Training strategy where cascade stages are trained sequentially, each stage using outputs from the previous stage as inputs for progressive refinement.
IoU-aware loss
Specialized loss function directly integrating the IoU metric into optimization, encouraging the model to produce bounding boxes with more precise localizations.
Progressive IoU thresholds
Increasing sequence of IoU thresholds used in Cascade R-CNN where each successive stage only trains on samples with IoU higher than the previous stage's threshold.
Contextual reasoning
Model's ability to consider spatial and contextual relationships between objects to improve detection accuracy, particularly important in complex scenes.