Human-Verified 3D Bounding Box Annotations for In-The-Wild Images

WildDet3D-Bench provides 9,256 human-verified 3D bounding box annotations across 2,196 images, forming a validation set for monocular 3D object detection. The 2,470 images are sourced from the COCO and LVIS datasets. This visualization data was created by AllenAI and was last updated in April 2026.

Use Cases

• Benchmark model performance on 3D bounding box predictions using the 9,256 human-verified annotations as ground truth.
• Analyze detection failures on diverse, in-the-wild scenes from the COCO and LVIS source images.
• Validate model robustness by testing on the 2,470-image validation set drawn from multiple established datasets.
• Visualize and compare predicted 3D boxes against the provided human-annotated 3D bounding boxes.

Strengths

• Contains 9,256 human-verified 3D bounding box annotations, ensuring high-quality ground truth.
• Curated from 2,470 images across established datasets (COCO, LVIS), providing scene diversity.
• Annotations are spread across 2,196 images, offering a substantial number of labeled samples.

Limitations

• Limited to 2,470 images, which is a relatively small sample size for training modern deep learning models.
• Serves only as a validation/benchmark set, not a training dataset.
• Specific annotation attributes (e.g., object classes, occlusion levels) and potential class imbalances are not detailed.

Provenance

Source

Images drawn from MS-COCO 2017 validation and LVIS v1.0 train sets.

Collection Method

Human-annotated 3D bounding boxes created for the WildDet3D-Bench benchmark.

Freshness

Last updated in April 2026.