EvaluateConfig

from stera import Evaluate, EvaluateConfig

cfg = EvaluateConfig(
    sync_target=95.0,
    hand_2_weight=0.0,
    depth_required=False,
)
Evaluate(session, config=cfg).show()

EvaluateConfig is a frozen-by-convention dataclass with three kinds of fields:

1. Color thresholds — (good_min, ok_min) pairs that drive the green / amber / red chips in the report.
2. Health-score deductions — target + weight pairs that drive the 0–100 score. Set any weight to 0 to disable that check.
3. Required-stream flags — booleans (depth_required, imu_required) that gate the corresponding *_missing_penalty. Both default to True.

Hand checks have three buckets:

• hand_1plus_* — frames with at least 1 hand (= any-hand).
• hand_2_* — frames with exactly 2 hands.
• frames_with_more_hands* — frames with strictly more than 2 hands. Metric only; no config knobs because it's a detection-error indicator rather than a quality signal.

Defaults at a glance

EvaluateConfig(
    # color thresholds (good_min, ok_min)
    health_thresholds       = (80.0, 60.0),
    depth_valid_thresholds  = (80.0, 50.0),
    sync_thresholds         = (90.0, 70.0),
    hand_any_thresholds     = (70.0, 30.0),
    hand_1plus_thresholds   = (40.0, 15.0),
    hand_2_thresholds       = (30.0, 10.0),
    imu_gravity_max_dev     = 0.5,        # m/s² Δ from 9.81 for green

    # required-stream flags (True = penalise when missing)
    depth_required          = True,
    imu_required            = True,

    # health-score deductions
    rgb_gap_max_penalty     = 10.0,
    depth_valid_target      = 80.0,
    depth_valid_weight      = 0.3,
    depth_missing_penalty   = 10.0,
    pose_missing_penalty    = 15.0,
    imu_missing_penalty     = 5.0,
    sync_target             = 90.0,
    sync_weight             = 0.1,
    hand_any_target         = 50.0,
    hand_any_weight         = 0.15,
    hand_1plus_target       = 30.0,
    hand_1plus_weight       = 0.05,
    hand_2_target           = 20.0,
    hand_2_weight           = 0.10,
    hand_missing_penalty    = 0.0,
)

Color thresholds

A value ≥ good_min renders green, ≥ ok_min amber, otherwise red. Higher is always better.

imu_gravity_max_dev (default 0.5 m/s²) is a single number, not a pair. The |gravity| chip in the IMU section is green when |9.81 − gravity_magnitude| ≤ imu_gravity_max_dev, amber otherwise.

Required-stream flags

Two booleans control whether a missing stream triggers its flat *_missing_penalty. Both default to True — the score will drop when these streams are absent. Flip to False to make the stream optional for the score (the corresponding sections in the report still render as "no data" either way).

Pose is mandatory unconditionally (the SDK assumes ARKit pose is always present for an MCAP from the Stera App). If you need to make pose optional, set pose_missing_penalty=0.

Health-score deductions

The full deduction logic lives in stera.eval.metrics.compute_health. See Health score for the formulae and a worked example.

Override recipes

Stricter sync

EvaluateConfig(sync_target=95.0, sync_weight=0.2)

Raises the bar to 95 % of RGB frames within 50 ms of their nearest depth/pose sample, and doubles the deduction so it dominates the score.

Single-handed task

EvaluateConfig(hand_2_weight=0.0, hand_2_thresholds=(0.0, 0.0))

Stops penalising low exactly-2-hand frame counts and removes the red colour on the "2 hands" chip. Useful for recordings where one hand is holding the camera.

Rig without depth / IMU

EvaluateConfig(depth_required=False, imu_required=False)

Marks both streams optional. A recording from a non-LiDAR phone or an external camera no longer takes the flat 10 + 5 hit for missing those topics.

Headless QC: only fail on missing streams

EvaluateConfig(
    depth_valid_weight=0.0,
    sync_weight=0.0,
    hand_any_weight=0.0,
    hand_1plus_weight=0.0,
    hand_2_weight=0.0,
    rgb_gap_max_penalty=0.0,
)

Disables every threshold-based check. Score only drops when a whole stream is missing (depth, pose, IMU), which is the bare minimum you'd want in a CI pipeline.

Require hand annotations

EvaluateConfig(hand_missing_penalty=20.0)

Run this after a detection pipeline that should have called session.add_hand_pose. If the buffer is empty (the loop never wrote any), the score drops by 20.

Inspecting the active config

Evaluate.compute() puts the config it ran with on the metrics dict, and the report dumps the same in Technical reference → Health score config:

metrics = Evaluate(session, config=my_cfg).compute()
assert metrics["config"] is my_cfg