NVIDIA-TAO TAO Analyze Gaps OD mAP
Use this skill to run TAO Data Services object-detection gap analysis. The skill compares ground-truth and inference annotations, computes per-image per-class TP/FP/FN/AP50 metrics, and identifies weak images where any class metric falls below its threshold. It does not run inference; upstream steps must produce the inference annotations first.
The container entrypoint is:
gap_analysis object_detection -e /absolute/path/to/object_detection.yaml
Inputs
Required spec fields:
| Field | Meaning |
|---|---|
ground_truth_ann_path |
KITTI label directory or COCO .json with ground-truth boxes. |
inference_ann_path |
KITTI label directory or COCO .json with model predictions. |
images_dir |
Root image directory. Establishes the full image universe including unannotated images. |
results_dir |
Output directory for all artifacts. |
kpi |
Identifier tag written to every output row. |
input_format |
kitti or coco. Must be declared explicitly; never inferred from the path. |
Common optional fields:
| Field | Default | Meaning |
|---|---|---|
iou_threshold |
0.5 |
IoU at or above which a prediction is accepted as a true positive. |
conf_threshold |
0.0 |
Predictions below this confidence are dropped before matching. |
min_area |
0 |
Boxes whose pixel area (w × h) is strictly below this value are discarded. |
class_mapping |
{} |
Maps raw annotation label strings to canonical class names. Absent labels are kept as-is. |
weak_thresholds |
{} |
Per-class thresholds as {class_name: {recall, precision, ap50}}. Absent keys fall back to the default_*_threshold values. Reference ITS defaults: car 0.99, bicycle 0.7, person 0.7 — a strict gate on the abundant, well-learned class and looser gates on the rare ones the loop exists to improve. |
default_recall_threshold |
0.5 |
Fallback recall threshold for classes not listed in weak_thresholds. |
default_precision_threshold |
0.0 |
Fallback precision threshold. Set to 0.0 to disable precision-based weak selection. |
default_ap50_threshold |
0.5 |
Fallback for classes absent from weak_thresholds. Set 0.0 so unlisted classes never mark an image weak — the reference filter had no fallback, and leaving TAO DS's 0.5 in place silently gates every class you did not list. |
Do not hand-write the spec. Copy the template and fill in the nulls — every
tuning value it already carries is the one this stage wants — then validate:
cp skills/data/tao-analyze-gaps-od-map/assets/default_object_detection.yaml "$SPEC"
# fill ground_truth_ann_path, inference_ann_path, images_dir, results_dir, kpi, input_format
python3 skills/data/tao-analyze-gaps-od-map/scripts/verify_object_detection_spec.py --spec "$SPEC"
verify rejects the spellings that fail — uppercase input_format, relative or
missing paths, a weak_thresholds entry that is a bare number rather than a
mapping — and reports every gated class plus the default_* fallbacks, so the
selection criteria behind a weak set are recoverable from the run's output. It
warns when a fallback is above zero, since that gates classes you did not list.
Quick Start
Run from the tao-skill-bank repo root.
Write the spec into the results directory. The run emits four artifacts and does not retain the spec, so a completed gap analysis otherwise cannot tell you which thresholds produced its weak set — and that weak set sizes the mining budget downstream. Keeping them together makes the selection criteria recoverable from the run alone.
RESULTS_DIR=/absolute/path/for/this/run # results_dir in the spec
SPEC="$RESULTS_DIR/object_detection.yaml" # spec lives beside its outputs
RUN_ROOT=/absolute/path/that/contains/annotations/images/and/results
GPU_COUNT=1
DS_IMAGE=nvcr.io/nvstaging/tao/tao-toolkit-ds:7.2.0-rc-36-multiarch # versions-key: images.tao_toolkit.data_services
docker run --rm --gpus "$GPU_COUNT" --ipc=host --network=host \
-v "$RUN_ROOT:$RUN_ROOT" \
-w "$RUN_ROOT" \
"$DS_IMAGE" \
gap_analysis object_detection -e "$SPEC"
Do not pass --user $(id -u):$(id -g); some TAO DS images call getpass.getuser() at startup and fail when the UID is not in /etc/passwd.
Preflight
- Verify Docker access:
docker info > /dev/null
- Resolve and pull the data-services image if needed:
DS_IMAGE=nvcr.io/nvstaging/tao/tao-toolkit-ds:7.2.0-rc-36-multiarch # versions-key: images.tao_toolkit.data_services
docker image inspect "$DS_IMAGE" > /dev/null || docker pull "$DS_IMAGE"
- Confirm
RUN_ROOTcontains the spec, both annotation sources, and the image directory. MountRUN_ROOTto the same absolute path inside Docker.
Outputs
| Artifact | Location | Contents |
|---|---|---|
| FP/FN box gaps | results_dir/box_gaps.parquet |
One row per unmatched box: kpi, image_id, filepath, class, gap_type (FP/FN), bbox, confidence, best_iou. |
| Per-image metrics | results_dir/image_metrics.parquet |
Per-image per-class: tp, fp, fn, precision, recall, ap50. |
| Weak images | results_dir/weak_images.parquet |
Images where any class metric falls below threshold: filepath, weak_classes, weak_recall, weak_precision, weak_ap50. Feed this into tao-mine-od-images. |
| Gap report | results_dir/gap_report.json |
FP/FN counts by type and class, plus run settings. |
All four artifacts are always written, even when no gaps are found.
Troubleshooting
The subtask object_detection requires -e/--experiment_spec_file: rerun with gap_analysis object_detection -e "$SPEC".
Input path not found inside Docker: use a RUN_ROOT mount where host and container paths are identical.
input_format error: set input_format: kitti or input_format: coco explicitly — it is never inferred from the path.
weak_images.parquet is empty: all class metrics are above their thresholds. Lower default_recall_threshold / default_ap50_threshold or add per-class entries to weak_thresholds.
Output directory not writable after Docker exits: the container writes as root. Chown back with docker run --rm -v "$RUN_ROOT:$RUN_ROOT" alpine chown -R "$(id -u):$(id -g)" "$RESULTS_DIR".