---

name: annotations-and-results description: "Guide SAHI annotation, prediction result, mask, COCO export, optional conversion, visualization, and file helper objects." disable-model-invocation: true metadata: disco-role: operating license: MIT

annotations-and-results

Use this sub-skill when a task is about SAHI's low-level data objects and result conversions: BoundingBox, Category, Mask, ObjectAnnotation, PredictionScore, ObjectPrediction, PredictionResult, COCO annotation/prediction dictionaries, optional FiftyOne or imantics conversion, visualization helpers, mask/bbox helpers, and lightweight file/import utilities.

Route away when the task owner is not object/result plumbing:

• Full prediction runs, detector setup, slicing parameters, batch prediction, CLI prediction, or video prediction: use ../sliced-inference/SKILL.md.
• Full COCO dataset creation, slicing, remapping, evaluation, YOLO export, or dataset-scale FiftyOne workflows: use ../dataset-tools/SKILL.md.
• NMS/NMM/GreedyNMM algorithm behavior or backend acceleration: use ../postprocess-backends/SKILL.md.

Start here

1. Read objects and exports before constructing or serializing annotation/prediction objects. It records constructor shapes, coordinate conventions, image-id handling, mask behavior, COCO JSON shapes, and optional FiftyOne/imantics conversions.
2. Read CV and file helpers when converting masks, drawing predictions, reading images, listing files, saving JSON, checking imports, or debugging OpenCV/Pillow color assumptions.
3. Read troubleshooting when boxes are invalid, category types fail, masks need full_shape, COCO masks disappear, optional conversion packages are missing, images load with wrong colors, or output paths are surprising.
4. For a dependency-light local sanity check, run python scripts/prediction_objects_smoke.py from this sub-skill directory. Add --visualize or --write-json only when you want temporary outputs.

Minimal decision checklist

• Decide which coordinate format is entering SAHI: core objects use [minx, miny, maxx, maxy]; COCO bboxes use [x, y, width, height].
• Keep image shapes as [height, width] for masks and full_shape; keep image IDs explicit when exporting per-image COCO predictions.
• Treat fiftyone and imantics as optional packages. Guard those conversions or give a fallback COCO JSON path.
• Use PredictionResult.export_visuals for simple PNG output; use lower-level visualization helpers when custom color or export format is required.
• Never let helper scripts download data, train models, require credentials, or write outside a caller-chosen or temporary output directory.