---

name: postprocess-backends description: "Guide SAHI NMS/NMM backend selection, match metrics, class-aware behavior, and direct postprocess APIs." disable-model-invocation: true metadata: disco-role: operating license: MIT

postprocess-backends

Use this sub-skill when the task is to choose or debug SAHI postprocessing backends and duplicate-removal behavior: set_postprocess_backend, get_postprocess_backend, resolve_backend, direct array calls such as nms, batched_nms, greedy_nmm, nmm, and object-list postprocessors such as NMSPostprocess, NMMPostprocess, and GreedyNMMPostprocess.

Do not use this sub-skill as the owner for detector loading, weights, or framework wrappers; route those to ../model-integrations/SKILL.md. For where to pass postprocess_type, postprocess_match_metric, postprocess_match_threshold, or class-agnostic options in prediction pipelines, route to ../sliced-inference/SKILL.md. For result object export, serialization, COCO JSON, or visualization payloads, route to ../annotations-and-results/SKILL.md.

Start here

1. Distinguish backend state before changing code:
   • get_postprocess_backend() returns the configured value, which can remain "auto".
   • resolve_backend() returns the concrete backend used by dispatch: "numpy", "numba", or "torchvision".
   • See Backend selection and acceleration limits.
2. Pick the postprocess operation:
   • Use NMS when duplicates should be discarded.
   • Use GreedyNMM when overlapping detections should be merged by direct highest-score neighbors.
   • Use NMM when transitive overlaps should also merge into the same keeper.
   • See Direct API and class reference.
3. Decide class behavior:
   • Class-agnostic mode lets all categories compete or merge together.
   • Class-aware mode runs independently per category; use batched_* direct functions or class_agnostic=False on postprocess classes.
4. Decide the match metric:
   • "IOU" is the default for similarly sized boxes.
   • "IOS" is often the correct choice for nested small objects, where a small box inside a larger duplicate has low IoU but high intersection-over-smaller.
5. If optional acceleration behaves inconsistently, force "numpy" first and run the bundled safe smoke script.
6. If the number of boxes is surprising, check Troubleshooting before changing model thresholds.

Safe local smoke

From this sub-skill directory, run:

python scripts/postprocess_backend_smoke.py --print-backend

The smoke script builds tiny in-memory arrays with columns [x1, y1, x2, y2, score, category_id], forces the numpy backend by default, exercises nms, batched_nms, greedy_nmm, and nmm, and asserts deterministic outputs. It does not download data, train models, use credentials, contact the network, or write files.

Boundaries and assumptions

• Backend selection is process-global and should be set at startup, before threaded inference or batched prediction loops.
• Optional acceleration is opportunistic: numba requires the optional package and has first-call JIT cost; torchvision requires compatible torch/torchvision and only auto-selects when CUDA or Apple MPS is available.
• Low-level direct APIs operate on numpy-compatible arrays and return indices or merge mappings, not serialized result objects.
• Prediction-pipeline parameter placement is intentionally excluded here; use ../sliced-inference/SKILL.md for get_sliced_prediction and CLI placement details.