---

name: weighted-boxes-fusion description: "Route Weighted-Boxes-Fusion tasks for 2D box ensembling, 1D span fusion, and 3D cuboid fusion with ensemble_boxes." disable-model-invocation: true metadata: disco-role: operating license: MIT

Weighted Boxes Fusion

Use this repo skill when a task asks you to merge predictions from multiple models with the ensemble_boxes package. The package covers three distinct geometry families:

• 2D detector boxes with WBF, NMS, Soft-NMS, and non-maximum weighted suppression.
• 1D spans or token intervals for NLP/NER-style fusion.
• 3D axis-aligned cuboids for volumetric or LiDAR-style detection.

The skill is split into focused sub-skills so future agents can route quickly by geometry and failure mode.

Install and verify

Install the public package from PyPI:

python -m pip install ensemble-boxes

The bundled smoke helper uses only the package's runtime dependencies; no benchmark data, GUI extras, or detector checkpoints are required.

Then run the root smoke helper:

python scripts/check_install.py --case all

That script imports ensemble_boxes and exercises tiny 2D, 1D, and 3D in-memory cases without downloading data or opening GUI windows.

Route map

sub-skills/box-ensembling/

Use this for 2D normalized boxes in [x1, y1, x2, y2] order.

Choose it for tasks like:

• merging YOLO, Faster R-CNN, DetectoRS, or other detector outputs;
• choosing between weighted_boxes_fusion, weighted_boxes_fusion_experimental, non_maximum_weighted, nms, soft_nms, and nms_method;
• tuning iou_thr, skip_box_thr, weights, conf_type, allows_overflow, sigma, or thresh;
• explaining clipped coordinates, swapped corners, zero-area boxes, or score inflation.

Read:

• sub-skills/box-ensembling/references/api-reference.md
• sub-skills/box-ensembling/references/workflows.md
• sub-skills/box-ensembling/references/troubleshooting.md
• sub-skills/box-ensembling/scripts/smoke_2d_ensembling.py

sub-skills/line-segment-fusion/

Use this for normalized 1D spans in [x1, x2] order.

Choose it for tasks like:

• fusing token spans or predictionstring intervals from several NLP models;
• mapping string classes to numeric labels before fusion and back afterward;
• tuning iou_thr, skip_box_thr, weights, conf_type, or allows_overflow for span ensembling.

Read:

• sub-skills/line-segment-fusion/references/api-reference.md
• sub-skills/line-segment-fusion/references/workflows.md
• sub-skills/line-segment-fusion/references/troubleshooting.md
• sub-skills/line-segment-fusion/scripts/smoke_1d_fusion.py

sub-skills/three-d-box-fusion/

Use this for normalized 3D cuboids in [x1, y1, z1, x2, y2, z2] order.

Choose it for tasks like:

• fusing LiDAR, medical, or other volumetric detector outputs;
• choosing between conf_type='avg' and conf_type='max';
• diagnosing axis-order mistakes, zero-volume boxes, or normalization problems.

Read:

• sub-skills/three-d-box-fusion/references/api-reference.md
• sub-skills/three-d-box-fusion/references/workflows.md
• sub-skills/three-d-box-fusion/references/troubleshooting.md
• sub-skills/three-d-box-fusion/scripts/smoke_3d_fusion.py

Shared guidance

• Keep all geometry inputs normalized unless a sub-skill explicitly shows the normalization step.
• Keep one outer list entry per model, even when a model has no predictions.
• If the task asks about benchmark reproduction, read references/benchmark-notes.md first. Those scripts need external data and optional packages, so they are not part of the default smoke path.
• If import or installation fails, start with references/troubleshooting.md before guessing at geometry bugs.
• If the skill seems stale for the current checkout, read references/repo-provenance.md and refresh the skill instead of reusing it blindly.
• For router integration metadata, consult references/repo-routing-metadata.json.

Minimal runtime check

The fastest package check is:

python -c "from ensemble_boxes import weighted_boxes_fusion; print(weighted_boxes_fusion)"

For a broader smoke, use scripts/check_install.py --case all.

What not to expect

• No training loop, model zoo, or detector architecture guidance.
• No built-in GUI visualization requirement.
• No default benchmark downloads.
• No GPU dependency for the package APIs covered here.