AI Glossary
The complete dictionary of Artificial Intelligence
Oriented Bounding Box
A form of object detection where the prediction box is defined by its center, width, height, and a rotation angle, allowing for a better fit around elongated or tilted objects.
Prediction Angle
The angular value, usually in radians or degrees, predicted by a model to orient a bounding box, essential for non-axis-aligned object detection.
Parametric Representation
A method for describing a rotating box by a set of parameters (x, y, w, h, θ) instead of the coordinates of its four vertices, optimizing loss calculations.
Angle Regression Loss
A specific cost function that penalizes the difference between the predicted angle and the ground truth angle of the bounding box, often based on L1 or L2 loss.
Periodicity Problem
The ambiguity where a box oriented by an angle θ and another by θ+π represent the same box, which complicates angle regression and requires specific encoding strategies.
Sine-Cosine Encoding
A technique to represent the angle of a rotating box using sin(θ) and cos(θ) values to avoid the discontinuity problem at the π/2 boundary.
Rotated IoU
An evaluation metric that calculates the Intersection over Union between two oriented bounding boxes, taking into account their respective rotations to measure detection accuracy.
Oriented Anchor
Pre-defined reference boxes with different sizes, aspect ratios, and angles, used by anchor-based models to more accurately predict rotating boxes.
Oriented Anchor-Free Detection
A detection approach that directly predicts the parameters of the rotating box from key points of the image, without using predefined anchor boxes.
Five-Parameter Regression
The process of simultaneously predicting the five parameters defining a rotating box: the center coordinates (x, y), width (w), height (h), and rotation angle (θ).
Rotated NMS
A variant of the Non-Maximum Suppression algorithm that calculates the overlap between oriented boxes using rotated IoU to eliminate redundant detections.
Angle Focusing Loss
An advanced loss function that gives more weight to samples with misclassified angle errors, improving the model's robustness for heavily inclined objects.
Multi-Orientation Detection
The ability of a model to detect objects with varied orientations within the same image, a key challenge for autonomous driving or aerial imaging systems.
Oriented Feature Aggregation
A technique where features extracted from a region of interest are aligned or transformed according to the predicted orientation before final classification.
Vertex Regression
An alternative to five-parameter regression, which involves directly predicting the coordinates of the four vertices of the rotating box, offering greater shape flexibility.
Skewness Loss
A loss function that penalizes predicted boxes whose orientation is incorrect relative to the main axis of the object, measuring the asymmetry of the prediction.
Angle Calibration
A post-processing or dedicated network layer to refine angle predictions to correct systematic model errors, often based on fine regression.