AI Glossary
The complete dictionary of Artificial Intelligence
Frustum PointNets
Neural network architecture that operates in a frustum of vision projected from a 2D bounding box to predict an accurate 3D bounding box.
Bird's-Eye-View (BEV) Projection
Technique for transforming sensor data (images, point clouds) into a map-like top-down representation to simplify object detection and localization.
Multi-View Fusion
Process of aggregating information from multiple viewpoints (cameras) to reconstruct a consistent and robust 3D scene or object.
Ray Casting
Algorithm that casts virtual rays from an observation point to determine intersections with objects in a 3D scene, fundamental for rendering and projection.
3D Bounding Box
Rectangular parallelepiped parameterized by its position, its dimensions (length, width, height) and its orientation (yaw, pitch, roll angles) to enclose a 3D object.
IoU 3D (Intersection over Union)
Evaluation metric for 3D detection, calculating the ratio between the intersection volume and the union volume of the predicted bounding box and the ground truth bounding box.
LiDAR (Light Detection and Ranging)
Active ranging technology that emits laser pulses and measures the return time of their reflection to generate precise 3D point clouds of the environment.
Monocular 3D Object Detection
Task consisting of estimating the 3D position, size, and orientation of an object from a single 2D image, relying on geometric and contextual cues.
Stereo Vision
Method that uses two or more calibrated cameras to simulate human vision and calculate depth by triangulation from image disparity.
PointNet
Pioneering deep neural network designed to directly process unstructured point clouds, learning symmetric functions over the set of points.
Range Image
2D representation of depth sensor data where each pixel stores the measured distance, similar to an image but with intensity replaced by range.
Anchor-based 3D Detection
Detection approach that predicts offsets from a predefined grid of 3D boxes (anchors) of different sizes and orientations in space.
Center-based 3D Detection
Detection paradigm that directly predicts the center of each object and its properties (size, orientation), simplifying post-processing and avoiding complex IoU calculations.