Class AnchorMatcher<T>

Namespace

AiDotNet.ComputerVision.Detection.Anchors

Assembly

AiDotNet.dll

Matches anchor boxes to ground truth boxes for training object detectors.

public class AnchorMatcher<T>

Type Parameters

T

The numeric type used for calculations.

Inheritance

object

AnchorMatcher<T>

Inherited Members

object.Equals(object)

object.Equals(object, object)

object.GetHashCode()

object.GetType()

object.MemberwiseClone()

object.ReferenceEquals(object, object)

object.ToString()

Remarks

For Beginners: During training, we need to assign each anchor box to either: - A ground truth box (positive sample): The anchor should predict this object - Background (negative sample): The anchor should predict "no object" - Ignore: The anchor is borderline and excluded from loss calculation

This matching process determines which anchors learn to detect which objects.

Constructors

AnchorMatcher()

Creates a new anchor matcher with default thresholds.

public AnchorMatcher()

AnchorMatcher(double, double, bool)

Creates a new anchor matcher with custom thresholds.

public AnchorMatcher(double positiveThreshold, double negativeThreshold, bool allowLowQualityMatches = true)

Parameters

positiveThreshold double

IoU threshold for positive matches.

negativeThreshold double

IoU threshold for negative matches.

allowLowQualityMatches bool

Whether to match best anchor per GT regardless of threshold.

Properties

AllowLowQualityMatches

Whether to allow low-quality matches (best anchor for each GT even if below threshold).

public bool AllowLowQualityMatches { get; }

Property Value

bool

NegativeThreshold

IoU threshold below which an anchor is considered negative.

public double NegativeThreshold { get; }

Property Value

double

PositiveThreshold

IoU threshold above which an anchor is considered positive.

public double PositiveThreshold { get; }

Property Value

double

Methods

Match(List<BoundingBox<T>>, List<BoundingBox<T>>)

Matches anchors to ground truth boxes.

public AnchorMatchResult<T> Match(List<BoundingBox<T>> anchors, List<BoundingBox<T>> groundTruth)

Parameters

anchors List<BoundingBox<T>>

List of anchor boxes.

groundTruth List<BoundingBox<T>>

List of ground truth boxes.

Returns

AnchorMatchResult<T>

Match result containing assignments for each anchor.

Remarks

For Beginners: This method looks at each anchor and decides: - If it overlaps enough with a ground truth box → match to that box - If it doesn't overlap much with any box → mark as background - If it's in between → ignore during training

MatchCenterBased(List<(double X, double Y)>, List<BoundingBox<T>>, List<int>)

Matches anchors using center-based assignment (used in FCOS, YOLOX).

public AnchorMatchResult<T> MatchCenterBased(List<(double X, double Y)> anchorCenters, List<BoundingBox<T>> groundTruth, List<int> anchorStrides)

Parameters

anchorCenters List<(double X, double Y)>

groundTruth List<BoundingBox<T>>

List of ground truth boxes.

anchorStrides List<int>

Returns

AnchorMatchResult<T>

Match result with center-based assignment.

Remarks

For Beginners: Center-based matching assigns an anchor to a GT box if the anchor's center point falls inside the GT box. This is used by anchor-free detectors like FCOS and YOLOX.

MatchSimOTA(List<BoundingBox<T>>, List<BoundingBox<T>>, List<BoundingBox<T>>, List<double>)

Matches anchors using SimOTA (used in YOLOX).

public AnchorMatchResult<T> MatchSimOTA(List<BoundingBox<T>> anchors, List<BoundingBox<T>> groundTruth, List<BoundingBox<T>> predictions, List<double> predScores)

Parameters

anchors List<BoundingBox<T>>

List of anchor boxes.

groundTruth List<BoundingBox<T>>

List of ground truth boxes.

predictions List<BoundingBox<T>>

Predicted boxes for cost calculation.

predScores List<double>

Predicted scores for cost calculation.

Returns

AnchorMatchResult<T>

Match result using optimal transport assignment.

Remarks

For Beginners: SimOTA (Simplified Optimal Transport Assignment) is an advanced matching strategy that considers both location and predicted confidence to make better assignments. It dynamically decides how many anchors each GT gets.