Detection knobs¶

For backends whose task_kind == detection (e.g. torchvision Faster R-CNN / RetinaNet / SSD), the pipeline switches to a per-box explanation loop that emits one ExplanationResult per detected box. Knobs live under the explainer's raitap.detection block; defaults apply when omitted.

The explainer's call.target MUST be 0 for detection runs (the ScalarDetectionWrapper exposes one scalar channel; target=0 selects it). call.target: auto_pred is rejected with a RaitapError because argmax over a 1-channel output always returns 0, which would mask real config errors.

Options

Name	Allowed	Default	Description
`raitap.detection.score_threshold`	`float`	`0.5`	Drop detections whose score is strictly below this before selecting boxes. Applies per-sample; samples with no detections above the threshold emit a warning and skip the K-loop.
`raitap.detection.max_boxes`	`int`	`5`	Cap K per sample after threshold filtering, by score descending. Smaller values keep CI / demo runs fast; larger values give more attribution coverage at the cost of one extra explainer call per box.
`raitap.detection.iou_threshold`	`float`	`0.5`	IoU threshold used by the `reference_match` target. Perturbed predictions must have at least this IoU with the original box (and a matching label) to count toward the explained scalar; below this the scalar smoothly degrades, preserving autograd for gradient explainers.

Examples

YAML

transparency:
  my_ig_explainer:
    _target_: "CaptumExplainer"
    algorithm: "IntegratedGradients"
    call:
      target: 0
    raitap:
      detection:
        score_threshold: 0.5
        max_boxes: 5
        iou_threshold: 0.5
    visualisers:
      - _target_: "DetectionImageVisualiser"

Python

from raitap.transparency import captum, detection_image

transparency = {
    "my_ig_explainer": captum(
        algorithm="IntegratedGradients",
        call={"target": 0},
        raitap={
            "detection": {
                "score_threshold": 0.5,
                "max_boxes": 5,
                "iou_threshold": 0.5,
            }
        },
        visualisers=[detection_image()],
    ),
}

CLI

uv run raitap transparency.my_ig_explainer.raitap.detection.max_boxes=3

pip

raitap transparency.my_ig_explainer.raitap.detection.max_boxes=3

Box labels and ground truth¶

Boxes are labelled in the report headings and on the thumbnail overlay; the per-box attribution figures stay title-less (a title would only duplicate them).

Predicted name. A box reads kite, not class 38, when a name source is available: torchvision detectors use their bundled categories automatically; otherwise set model.class_names (id-ordered, index 0 first), which takes precedence. With neither, boxes fall back to class <id>.

True label. Set data.labels.source to match each box to the ground-truth object it overlaps most, shown as gt: <name> (IoU <value>). The match is by overlap alone, so disagreements surface (pred: dog 0.92 | gt: cat (IoU 0.71)); a box overlapping no labelled object reads gt: no match, and with no ground truth the gt: clause is omitted. Matching reuses raitap.detection.iou_threshold.