OpenBox aligns 2D vision-foundation instance cues with 3D point clouds to produce more accurate and cheaper 3D box annotations without self-training across three autonomous-driving datasets

It proposes OpenBox, a two-stage automatic 3D bounding box annotation pipeline whose core is to use a 2D vision foundation model for instance-level cues and to associate those cues with the corresponding 3D point clouds via cross-modal instance alignment. Whereas existing approaches rely on multiple self-training iterations to refine annotations, OpenBox moves the annotation source to 2D foundation-model instance cues plus cross-modal alignment, so the pipeline design no longer requires self-training. The abstract explicitly describes the two-stage structure and the cross-modal instance alignment mechanism and states that results require no self-training; the alignment algorithm details, network architecture, and hyperparameters are not expanded in the given text.

The second stage classifies instances by rigidity and motion state and uses class-specific size statistics to generate adaptive boxes, rather than applying one uniform box-generation scheme to all objects. Against the existing practice of annotating 3D bounding boxes uniformly while ignoring objects' physical states, OpenBox explicitly introduces physical state (rigidity, motion) as a classification dimension and lets box size statistics vary by class. The abstract directly states the classification by rigidity and motion state and the class-specific size statistics; the concrete statistics per class, the classifier form, and thresholds are not given in the text.

Experiments on the Waymo Open Dataset, the Lyft Level 5 Perception dataset, and the nuScenes dataset show improvements in both accuracy and efficiency over baselines. Reporting gains along both accuracy and efficiency echoes the problem framing in the abstract about suboptimal quality and substantial computational overhead in prior methods. Evidence comes from experiments on three named autonomous-driving datasets; the given text provides no specific metric values, baseline names, ablation settings, or statistical significance information.

The work targets 3D object detection annotation in autonomous-driving scenes, suited to data-production pipelines that aim to cut manual annotation cost and need to recognize unseen categories; its setting is to generate 3D bounding boxes without self-training by leveraging a 2D vision foundation model and cross-modal instance alignment, with adaptive generation driven by rigidity and motion state and class-specific size statistics. The validation scope stated in the abstract is the Waymo Open Dataset, the Lyft Level 5 Perception dataset, and the nuScenes dataset, so directly reusable settings are the sensor and road environments those datasets represent; transferring to other sensor configurations, other object categories, or other domains requires re-evaluating the alignment and size statistics.

The given text is abstract-level information and does not include specific accuracy and efficiency values, baseline names, ablation studies, failure-case analyses, or how the class-specific size statistics are constructed, so the magnitude of improvement and which categories or scenes benefit most cannot be judged. How cross-modal instance alignment behaves under occlusion, sparse point clouds, or missing 2D detections, and the criteria for the rigidity and motion-state classification, are open questions that require the original paper. In addition, the abstract mentions a project page but the text does not provide verifiable link content, so implementation details and data splits await the original text.