Agent skill · nvidia
tao-analyze-detection-kpi
Run TAO Data Services KPI analysis for object detection, comparing inference annotations against ground truth to compute per-class TP/FP/FN/TN, precision, recall, accuracy, and AP at a fixed IoU. Use when an object detection workflow needs per-class mAP reported after inference, or when the user asks to "run KPI analyze", "compute detection mAP", or "score my OD predictions against ground truth".
What it needs
About 6k tokens when loaded.
What this skill does
TAO Analyze Detection KPI Use this skill to run TAO Data Services KPI analysis for object detection. The skill compares inference annotations against ground truth over one or more KPI sources and writes a per-class metrics CSV. It does not run inference; an upstream step must produce the inference annotations first. The container entrypoint is: Inputs The user provides either a finished spec or the paths to fill into the template. Required spec fields: Field Meaning --- --- data.inputformat KITTI or COCO. Uppercase — see Pitfalls. data.kpisources List of sources. Each entry requires imagedir, groundtruthannpath, and inferenceannpath; all three are asserted at startup. data.mapping Path to a class-mapping YAML: a list of single-key dicts whose value is a LIST of aliases. See assets/examplemapping.yaml — a bare string here silently zeroes every metric. resultsdir Output directory for kpicalc.csv. Common optional fields. The Default column is what TAO DS uses when the field is absent; assets/defaultkpianalyze.yaml already carries the recommended value for each, so filling the template needs none of them changed: Field Default Meaning --- ---: --- kpi.iouthreshold 0.5 IoU at or above which a prediction counts as a true positive. kpi.confthreshold 0.5 Predictions below this are dropped. The template uses 0.0, which keeps the whole PR curve so a threshold can be swept afterwards without re-running inference. On the pinned image that is safe: unmatched ground truth carries a -1.0 sentinel and lands in FN at any threshold. On a build predating that fix, 0.0 scored every missed box as a true positive — TP became the ground-truth count and FN was always 0 — so use a small positive value there. kpi.numrecallpoints 11 Recall points for the interpolated PR curve. The template keeps 101 (COCO-style), matching the reference ITS pipeline. 11 selects VOC-style sampling and reports different numbers for the same detections. kpi.ignoresqwidth 0 Boxes narrower than this are ignored. …
How to use it
Reference it in AdaL, Claude Code, Cursor or any coding agent — nothing to install:
@skills nvidia/tao-analyze-detection-kpi