3D Bounding Box Annotations for In-the-Wild Images

WildDet3D-Data provides 3D bounding box annotations for images sourced from major detection datasets like COCO and LVIS. The training split with human-reviewed annotations contains 102,979 images and 229,934 annotations across 11,879 categories. It was created by Allen Institute for AI (allenai) and was updated in April 2026.

Use Cases

• Train 3D object detection models using the provided 3D bounding box annotations on in-the-wild images.
• Benchmark model performance on the 'Human' split with 229,934 verified annotations versus the 'Essential' split with 412,711 VLM-qualified annotations.
• Study small object detection by leveraging the VLM-qualified annotations for small objects in the 'Essential' split.
• Analyze category coverage across the 11,879 to 12,064 annotated object categories for model generalization.

Strengths

• Contains 102,979 in-the-wild images from established sources like COCO and LVIS.
• Human-reviewed 'Train' split provides 229,934 annotations with high reliability.
• Offers a large annotation set of 412,711 entries in the 'Essential' split, combining human and VLM-qualified labels.

Limitations

• Specific image sources and collection methodology details are not fully described in the provided input.
• The 'Synthetic' split details and validation/test set statistics are not provided, limiting completeness assessment.
• Potential for label noise or bias inherent in the original source datasets (COCO, LVIS, Objects365, V3Det) may persist.

Provenance

Source

Annotations are derived from images in COCO, LVIS, Objects365, and V3Det datasets.

Collection Method

Annotations are a combination of human-reviewed labels and labels qualified by a Vision-Language Model (VLM).

Freshness

Last updated on 2026-04-07, indicating recent maintenance.