oa::FnDetection

FnDetection namespace in the OA Vision public surface.

NsC++ public headers

Public Functions

Matrix oa::FnDetection::binaryMaskCounts(const Matrix & inPredicted, const Matrix & inTarget)

Matrix oa::FnDetection::boxIou(const Matrix & inA, const Matrix & inB)

Matrix oa::FnDetection::confusionMatrix(const Matrix & inPredicted, const Matrix & inTarget, oa::I32 inClassCount)

DetectionMetricsResult oa::FnDetection::evaluate(const Matrix & inPredictedBoxes, const Matrix & inPredictedScores, const Matrix & inPredictedClasses, const Matrix & inPredictedImageIds, const Matrix & inTargetBoxes, const Matrix & inTargetClasses, const Matrix & inTargetImageIds, const Matrix & inIouThresholds, oa::I32 inClassCount, oa::F32 inScoreThreshold = 0.0F)

SegmentationMetricsResult oa::FnDetection::evaluateSegmentation(const Matrix & inPredicted, const Matrix & inTarget, oa::I32 inClassCount)

NmsResult oa::FnDetection::nms(const Matrix & inBoxes, const Matrix & inScores, const Matrix & inClasses, const NmsConfig & inConfig = {})

Function Documentation

Matrix oa::FnDetection::binaryMaskCounts(
	const Matrix & inPredicted,
	const Matrix & inTarget
)

Binary mask counts [true-positive, false-positive, false-negative, true-negative]. Inputs are equally-shaped UInt8 masks.

Parameters

inPredicted

const Matrix &

—

inTarget

const Matrix &

—

Returns

Matrix

The declared return value.

Matrix oa::FnDetection::boxIou(
	const Matrix & inA,
	const Matrix & inB
)

Pairwise IoU for FP32 center-x/center-y/width/height boxes. inA [N,4], inB [M,4] -> FP32 [N,M]. Coordinates may be normalized or pixel-valued, but widths and heights must be non-negative.

Parameters

inA

const Matrix &

—

inB

const Matrix &

—

Returns

Matrix

The declared return value.

Matrix oa::FnDetection::confusionMatrix(
	const Matrix & inPredicted,
	const Matrix & inTarget,
	oa::I32 inClassCount
)

Classification confusion matrix. rows are target classes and columns are predicted classes: Int32 [N], Int32 [N] -> UInt32 [C,C].

Parameters

inPredicted

const Matrix &

—

inTarget

const Matrix &

—

inClassCount

oa::I32

—

Returns

Matrix

The declared return value.

DetectionMetricsResult oa::FnDetection::evaluate(
	const Matrix & inPredictedBoxes,
	const Matrix & inPredictedScores,
	const Matrix & inPredictedClasses,
	const Matrix & inPredictedImageIds,
	const Matrix & inTargetBoxes,
	const Matrix & inTargetClasses,
	const Matrix & inTargetImageIds,
	const Matrix & inIouThresholds,
	oa::I32 inClassCount,
	oa::F32 inScoreThreshold = 0.0F
)

Dataset-level object-detection evaluation. Predictions are FP32 boxes [P,4], FP32 scores [P], Int32 classes [P] and Int32 image IDs [P]. targets are FP32 boxes [G,4], Int32 classes [G] and Int32 image IDs [G]. IoU thresholds are FP32 [T]. Matching is greedy by descending score, class-aware and constrained to the same image. classes without targets do not contribute to mAP. All outputs and internal scratch stay GPU-resident.

Parameters

inPredictedBoxes

const Matrix &

—

inPredictedScores

const Matrix &

—

inPredictedClasses

const Matrix &

—

inPredictedImageIds

const Matrix &

—

inTargetBoxes

const Matrix &

—

inTargetClasses

const Matrix &

—

inTargetImageIds

const Matrix &

—

inIouThresholds

const Matrix &

—

inClassCount

oa::I32

—

inScoreThreshold

oa::F32

Default: 0.0F

Returns

DetectionMetricsResult

The declared return value.

SegmentationMetricsResult oa::FnDetection::evaluateSegmentation(
	const Matrix & inPredicted,
	const Matrix & inTarget,
	oa::I32 inClassCount
)

Multiclass semantic-segmentation metrics over flattened Int32 label matrices. labels outside [0,classCount) are ignored by the confusion accumulator, which provides the ordinary ignore-index behavior.

Parameters

inPredicted

const Matrix &

—

inTarget

const Matrix &

—

inClassCount

oa::I32

—

Returns

SegmentationMetricsResult

The declared return value.

NmsResult oa::FnDetection::nms(
	const Matrix & inBoxes,
	const Matrix & inScores,
	const Matrix & inClasses,
	const NmsConfig & inConfig = {}
)

Deterministic class-aware NMS. boxes are FP32 [N,4] cx/cy/w/h, scores FP32 [N], and classes Int32 [N]. The implementation is GPU-resident and records as one graph node; it never sorts or compacts on the host.

Parameters

inBoxes

const Matrix &

—

inScores

const Matrix &

—

inClasses

const Matrix &

—

inConfig

const NmsConfig &

Default: {}

Returns

NmsResult

The declared return value.