How-to guides

Task-oriented recipes for users who already know what they want to accomplish.

• Submit detections as numpy arrays or DLPack tensors — skip JSON for tight model→evaluator loops.
• Evaluate on a background thread — BackgroundEvaluator for training loops where the kernel measurably stalls the main thread.
• Evaluate with boundary IoU — the Boundary() kernel and its dilation knob.
• Evaluate model calibration — ECE / MCE / reliability table for the detection family (ADR-0018), including the StreamingSnapshot pairing for training-loop folds.
• Evaluate with vernier eval — the static CLI binary for CI pipelines without a Python interpreter.
• Configure the evaluator — the entry-point survey of iou / parity_mode / max_dets / use_cats / cast_inputs and how they compose.
• Custom evaluation grids — the ADR-0040 axes (iou_thresholds, recall_thresholds, area_ranges) and the evaluate_tables route.
• Distributed evaluation across ranks — rank-local + gather across instance / semantic / panoptic.
• Evaluate keypoints with OKS — Keypoints() with per-category sigmas.
• Per-class metrics broken down by manifest slice — the recipe for the tables= + manifest= cross product (deliberate non-feature; the kwargs reject in combination).
• Use result tables — the per-image / per-class / per-detection / per-pair polars DataFrames.
• Slice an evaluation by scenario — manifest-driven (axis, value) cells, joint cells via cross_axes, and vernier aggregate for cross-run corruption tables (ADR-0046).