Multi-View 3D Object Detection

The field uses images from multiple viewpoints to estimate 3D object locations, sizes, and classes, supporting applications such as autonomous driving and indoor mapping. Common approaches lift image features into bird’s-eye-view representations, use query-based transformers to combine camera views, or exploit depth, geometric constraints, and temporal information. Methods are evaluated on driving benchmarks such as nuScenes and Waymo and indoor datasets such as ScanNet and ARKitScenes. A central challenge is maintaining accurate, efficient detection when camera layouts, viewing conditions, or domains change, while extending recognition to unfamiliar object categories.