Three-dimensional object detection extends traditional 2D bounding boxes into the physical world by predicting the location, size, and orientation of objects in 3D space. Instead of flat rectangular boxes on an image, the output is a set of 3D cuboids defined by center coordinates (x, y, z), dimensions (length, width, height), and a heading angle. This is essential for any system that needs to understand where objects actually are in the real world, not just where they appear in a camera frame.

Input data for 3D detection typically comes from LiDAR point clouds, depth cameras, stereo camera pairs, or a fusion of LiDAR and camera feeds. Point-cloud methods like PointPillars and CenterPoint convert raw 3D points into structured representations (pillars or voxels) and apply convolutional detection heads. Camera-only approaches like BEVDet and BEVFormer project image features into a bird's-eye-view representation and predict 3D boxes without any depth sensor. Fusion methods combine the rich texture from cameras with the precise depth from LiDAR to get the best of both.

Autonomous driving is the primary application, where every vehicle, pedestrian, and cyclist must be localized in 3D for safe path planning. Robotics, warehouse automation, and augmented reality also depend on 3D detection. Evaluation uses metrics like 3D IoU (intersection over union of cuboids) and BEV AP (average precision in bird's-eye view), which are stricter than their 2D counterparts because small errors in depth estimation cause large drops in 3D overlap.