benchmarks.hota
module unitrack.benchmarks.hotaHOTA model-benchmark harness.
Pairs open-weights HuggingFace models with the unitrack tracker and the
evaluators HOTA metric. See ``benchmarks/hota/README.md``.
Members
module
- datasetsDataset adapters. The first is cityscapes-dvps from a local LMDB.
- learned_modulesLearned MOTR-style appearance-filter modules.
- metricWrap the evaluators panoptic_tracking() preset as a streaming runner.
- modelsHuggingFace panoptic-segmentation model adapters.
- protocolsStructural protocols for the three harness plug points.
- renderRender tracker output into an evaluators-compatible panoptic map.
- reportMarkdown + JSON writers for benchmark results.
- runnerSweep models x sequences, streaming predictions into the HOTA metric.
- trackerTracker factories (mask-IoU, cosine, cascade, Kalman) + a registry.
- train_learnedTrain the learned MOTR-style appearance filter and write its checkpoint.
- typesPlain data carriers exchanged between the harness plug points.
type
- BenchmarkResultOne ``(model, tracker)`` combo's aggregated metric output plus timing.
- BenchmarkRunnerRun each model over every sequence and aggregate HOTA per model.
- CityscapesDVPSDatasetReads ``cityscapes-dvps.{split}.lmdb`` (keys ``seq/frame/{image,panoptic}``).
- DatasetAdapterA source of ground-truth sequences plus its panoptic label space.
- FramePredictionOne frame of model output: per-instance thing masks + semantics.
- HFPanopticAdapterLoads a HF panoptic model and emits per-frame thing instances.
- ModelAdapterA HuggingFace model that emits one ``FramePrediction`` per image.
- PanopticMetricRunnerStream (gt, pred) panoptic maps into HOTA/CLEAR/Identity and aggregate.
- SequenceSampleOne ground-truth sequence: an id, length, and a frame iterator.
- TrackerFactoryA factory returning a fresh ``unitrack.MultiStream`` for one sequence.
- TrackRemapStable, per-sequence ``(semantic_class, track_id) -> 1-based instance`` map.
function
- build_mask_trackerBuild a single-stage mask-IoU tracker over a fixed ``(height, width)``.
- build_modelConstruct an :class:`HFPanopticAdapter` for a registry ``key``.
- default_tracker_factoryReturn a :class:`TrackerFactory` building a fresh mask-IoU ``MultiStream``.
- ids_per_detectionRecover per-detection track ids in detection order (``-1`` if gated out).
- render_pred_panopticPaint thing instances as ``semantic * offset + dense_instance``.
constant
Source: unitrack/benchmarks/hota/__init__.py:1