Conformal Boxes

What It Is

Conformal Boxes is this library's name for conformal bounding-box uncertainty around object detections.
The primary method here is "Adaptive Bounding Box Uncertainties via Two-Step Conformal Prediction."
The paper was an ECCV 2024 oral.
It provides post-hoc uncertainty intervals for object bounding boxes with conformal coverage guarantees.
The method targets multi-object detection, including autonomous-driving-style safety applications.
It is an uncertainty quantification method, not a detector architecture by itself.

Split conformal prediction turns calibration-set errors into prediction intervals for new samples.
A detector predicts class labels and bounding boxes.
Standard box conformalization can fail when box regression depends on the predicted class.
The two-step approach first accounts for class-label uncertainty.
It then propagates that label uncertainty into bounding-box intervals.
The method broadens coverage guarantees to include incorrectly classified objects.
Ensemble and quantile-regression variants make intervals adaptive to object size.

Inputs are a trained object detector, a held-out calibration set, detector class scores, predicted boxes, and ground-truth boxes.
The calibration set must be exchangeable with the deployment distribution for formal guarantees to apply.
Intermediate outputs are nonconformity scores and quantile thresholds.
Final outputs are intervals around box coordinates or shaded box uncertainty regions.
The method can produce two-sided intervals that contain the true box with target coverage.
It does not change the detector's raw box prediction unless integrated into a downstream decision rule.
It currently applies most directly to 2D bounding-box localization as reported in the paper.

Step one constructs a prediction set or uncertainty treatment for object class labels.
Step two calibrates coordinate-wise box intervals conditional on class uncertainty.
Per-class calibration is used when the class is known or included in the class prediction set.
Adaptive variants scale interval width based on uncertainty estimates from ensembles or quantile regression.
The method is model-agnostic and can wrap black-box detectors.
Runtime overhead is small because conformal thresholds are learned offline.
The resulting intervals can be passed to planners as conservative occupied regions.

The detector is trained normally before conformal calibration.
Calibration uses held-out labeled data and a chosen error rate alpha.
Evaluation checks empirical coverage against the target level, such as 1 - alpha.
The ECCV paper validates on real-world datasets for 2D bounding-box localization.
The paper reports that desired coverage levels are satisfied with actionably tight intervals.
It studies balanced coverage across object sizes, not only average coverage.
The primary metrics are coverage, interval tightness, and size-conditioned behavior.

Provides distribution-free uncertainty guarantees under exchangeability assumptions.
Can be applied post-hoc to existing detectors.
Makes detection uncertainty spatially explicit for planning and safety logic.
The two-step design handles class-conditioned box predictions better than naive box-only conformalization.
Adaptive intervals avoid making every object pay for worst-case uncertainty.
It is easier to audit than opaque neural uncertainty scores.

Guarantees depend on calibration data matching deployment data.
Coverage can fail under domain shift, new object classes, different cameras, weather, or annotation changes.
The method addresses localization uncertainty for detected objects, not missed detections.
It is reported for 2D boxes; 3D boxes, BEV footprints, and occupancy require additional adaptation.
Box intervals can become too wide for planning if the detector is weak or data are noisy.
Association between predictions and ground truth can affect calibration scores in crowded scenes.
It does not identify why a detector is uncertain.

Conformal box intervals are useful for conservative planning around pedestrians, ground crew, baggage carts, and service vehicles.
They can inflate occupied regions when camera detections are uncertain under glare, darkness, or occlusion.
Airside systems need 3D or BEV uncertainty, so 2D conformal boxes are only a starting point.
Calibration data must come from the airport ODD, including night operations, rain, snow, jet blast, and de-icing.
Missed object risk must be handled separately with recall testing and sensor redundancy.
The planner should consume conformal intervals as safety margins, not as proof that all hazards were detected.

Keep a dedicated calibration split that is never used for detector training or model selection.
Recalibrate intervals whenever cameras, labels, detector architecture, or ODD distribution changes.
Report empirical coverage by class, size, range, occlusion, lighting, and weather.
For 3D detection, extend scores to center, size, yaw, and BEV footprint intervals before deployment use.
Combine conformal intervals with track-level temporal smoothing to avoid frame-by-frame margin flicker.
Treat low coverage on rare airside classes as a safety blocker, even if aggregate coverage is acceptable.