TemporalFix
TemporalFix stabilizes and repairs frame-level detections without coupling an application to a detector or full tracking framework. It accepts validated NumPy data and returns boxes with track IDs, lifecycle state, provenance, and bounded heuristic uncertainty.
detector -> Detections -> global association -> temporal stabilization
-> provenance-labelled output
Install the dependency-light core:
Start with the five-minute quick start, then review the configuration, public API, and limitations. The release candidate has no published performance or accuracy claim; its benchmark tools record evidence for the machine and data on which they run.
What it does
- deterministic global IoU association with optional class gating;
- none, EMA, or constant-velocity Kalman box smoothing;
- confidence smoothing and decay through short gaps;
- class evidence voting and switch diagnostics;
- false-positive confirmation and bounded gap recovery;
- independent state for multiple streams;
- safe YAML configuration and inspectable presets.
What it does not do
TemporalFix is not appearance-based tracking or long-term re-identification. Its uncertainty is an interpretable lifecycle heuristic, not a calibrated error probability. Detector execution remains outside the package.