Bird's-Eye-View Representation

Bird's-eye-view (BEV) representation converts camera, LiDAR, radar, or combined sensor observations into a top-down spatial feature map, making it easier to reason about road layouts, objects, and motion. Common approaches include geometric projection and inverse perspective mapping, learned view-transform networks with attention, height-aware or probabilistic encodings, and fusion across sensors or time. BEV features support tasks such as 3D detection, occupancy and semantic mapping, localization, tracking, and driving or robot navigation. A central challenge is balancing accurate geometry and robustness to calibration errors, occlusion, and sensor failures against the latency and memory costs of projection and dense representations; benchmarks increasingly test deployment and downstream behavior as well as perception accuracy.