Introducing WildDet3D: Open-world 3D detection from a single image

Allen Institute (AI2) · 3 months ago

Researchers released WildDet3D, an open-source model for 3D object detection from single images that accepts multiple prompt types including text queries and point selections. The model achieves 34.2 Average Precision on the Omni3D benchmark with text prompts, a 5.8-point improvement over the previous best method, and reaches 22.6 AP on a new benchmark spanning over 700 object categories. The release includes over 1 million training images with 3.7 million 3D annotations, enabling broader applications in autonomous vehicles, robotics, and AR systems that previously required narrow domain-specific approaches.