---

name: three-d-box-fusion description: "Fuse normalized 3D cuboid detections with weighted_boxes_fusion_3d." disable-model-invocation: true metadata: disco-role: operating license: MIT

three-d-box-fusion

Use this sub-skill when you need to combine normalized 3D axis-aligned cuboids from multiple detectors with weighted_boxes_fusion_3d.

Use for

• Fusing LiDAR, medical, or volumetric detector boxes that are already in a common normalized scene range.
• Choosing between conf_type='avg' and conf_type='max' for 3D fusion.
• Checking the returned boxes, scores, and labels arrays.
• Confirming coordinate order, weight handling, and threshold behavior for 3D WBF.

Do not use for

• 2D box ensembling, NMS, Soft-NMS, or non-maximum weighted fusion.
• 1D interval or span fusion.
• Dataset-specific coordinate conversion or visualization workflows.
• Any 3D NMS/NMW helper: this package does not provide one.

Expected inputs

• boxes_list: one list per model, each box in [x1, y1, z1, x2, y2, z2] order.
• scores_list: one confidence score list per model.
• labels_list: one numeric label list per model.
• Optional weights, iou_thr, skip_box_thr, conf_type, and allows_overflow.

Output contract

• boxes: fused boxes as an (N, 6) NumPy array.
• scores: fused confidences as an (N,) NumPy array.
• labels: numeric labels as an (N,) NumPy array.
• Outputs are sorted by confidence descending.

Routing hints

• If the task is 2D, route to the sibling 2D sub-skill.
• If the task is 1D, route to the sibling 1D sub-skill.
• If the task mentions metric LiDAR boxes, normalize them to a common [0, 1] scene range first, then fuse.
• If the task asks for unsupported 3D confidence names such as absent_model_aware_avg, fall back to avg and explain that 3D WBF accepts only avg and max.

Bundled guidance

• API reference
• Workflows
• Troubleshooting
• Smoke script