Unitrack
module

states

module unitrack.states

State protocols and :class:`~unitrack.states.State` recipes for tracklet fields.

Members

module

  • baseState, Process, Observation, and Initializer protocols.
  • directionalvon Mises-Fisher directional filter for unit-norm embedding states.
  • emaEMA-family :class:`~unitrack.states.Process` and :class:`Observation` primitives.
  • galleryGallery (feature-bank) state — keep a ring buffer of recent embeddings.
  • identityIdentity ``Process``, ``Replace``, and simple ``Initializer`` recipes.
  • kalmanKalman filter state family for unitrack.
  • learnedLearned propagation hooks — a slot for a MOTR-style recurrent update.
  • softDifferentiable :class:`Observation` companions used under ``differentiable=True``.

type

  • ConstantInitializer:class:`Initializer` that fills the schema-shaped buffer with a constant.
  • EMADecayExponential decay of a tracklet field toward ``anchor``.
  • EMAFuseExponential-moving-average update of a tracklet field.
  • EMATrackPredict-step companion to :class:`EMAFuse`.
  • EnsembleInitializer:class:`~unitrack.states.Initializer` spawning an ensemble around a measurement.
  • EnsembleProcessRandom-walk predict step: multiplicative covariance inflation.
  • EnsembleUpdateDeterministic ETKF measurement update for matched tracklet-detection pairs.
  • EyeInitializer:class:`Initializer` that emits a per-tracklet scaled identity matrix.
  • FromDetectionField:class:`Initializer` that copies a named field from new detections.
  • GalleryAppendUpdate step: push each matched detection's embedding into the ring buffer.
  • GalleryInitializer:class:`~unitrack.states.Initializer` seeding a gallery with one embedding.
  • IdentityNo-op :class:`~unitrack.states.Process` for fields that need no prediction step.
  • InformationProcessRandom-walk predict step in information form.
  • InformationUpdateAdditive measurement update in information form.
  • InitializerProduce field-shaped initial values for newly-promoted tracklets.
  • KalmanBBoxConstant-velocity bounding-box Kalman process.
  • KalmanCentroid2DConstant-velocity 2-D centroid Kalman process.
  • KalmanCentroid3DConstant-velocity 3-D centroid Kalman process.
  • KalmanLinearLinear-Gaussian Kalman predict step.
  • KalmanUpdateKalman measurement update (Joseph form) for matched tracklet-detection pairs.
  • LearnedObservationUpdate step that fuses matched detections with a learned module.
  • LearnedProcessPredict step that applies a learned module to every tracklet's field.
  • NoopObservation:class:`Observation` that leaves the snapshot unchanged.
  • NoopProcess:class:`~unitrack.states.Process` that leaves the snapshot unchanged.
  • NormalizedFromDetectionField:class:`Initializer` copying a detection field, L2-normalised along the last axis.
  • ObservationUpdate-step protocol: fuse detection measurements into tracklets.
  • PadZerosInitializer:class:`Initializer` that copies a detection field and zero-pads to ``full_dim``.
  • ProcessPredict-step protocol: advance a tracklet field by one time step.
  • ReplaceHard-replace :class:`Observation`: matched tracklets adopt the detection value.
  • ScaledFromDetectionField:class:`Initializer` copying a detection field scaled by a constant.
  • SoftReplaceDifferentiable :class:`~unitrack.states.Replace` driven by a transport plan.
  • StateA named field on the tracklet snapshot with predict/update/init logic.
  • VonMisesFisherDecayPredict step for a von Mises-Fisher embedding state.
  • VonMisesFisherUpdateConjugate von Mises-Fisher update for matched tracklet-detection pairs.
  • WeightedFuseScore-aware blend that uses a per-detection score as the blend weight.
  • ZerosInitializer:class:`Initializer` that fills the schema-shaped buffer with zeros.

function

Source: unitrack/states/__init__.py:1

merge

base