roboflow/trackers

BoTSORTTracklet and McByteTracklet already share ~250 near-identical lines (_configure_initial_noise, _build_process_noise, _build_measurement_noise, _set_scale_aware_noise, both _refresh_*_noise_from_state, all _clamp_* static methods, _SIGMA_P/_SIGMA_V/_SIGMA_M). PR #578 added a 5-line _diagonal_matrix helper duplicated verbatim across both files, on top of the existing duplication. Extract to a shared base class or mixin; fold _diagonal_matrix into that work. Shared home options: utils/matrix.py, module-internal privacy utils/_matrix.py, exposing a public diagonal_matrix — a third consumer already exists at src/trackers/utils/motion_models.py:236 Constraint: do not add to src/trackers/__init__.py — that surface is the stable public API per repo convention. Source: /oss:review PR #578 report, .reports/review/2026-09-10T11-04-46Z/review-report.md, finding A1.

Description MOT preprocessing follows TrackEval, but the ground-truth class sets are hardcoded to MOT17. TrackEval branches on the benchmark and we do not: # trackeval/datasets/mot_challenge_2d_box.py distractor_class_names = ['person_on_vehicle', 'static_person', 'distractor', 'reflection'] if self.benchmark == 'MOT20': distractor_class_names.append('non_mot_vehicle') # src/trackers/io/mot.py:31 # TODO(MOT20): MOT20 additionally treats non_mot_vehicle (6) as a distractor; # thread a benchmark parameter through _prepare_mot_sequence to support it. _DISTRACTOR_CLASSES = (2, 7, 8, 12) Ground-truth rows fall into three buckets. Scored: pedestrian (1), where a matched detection is a TP. Distractor: dropped before the metrics, no penalty. Neither: a detection on top of it counts as a false positive. MOT20 puts non_mot_vehicle (6) in the second bucket and we put it in the third, so on MOT20 a tracker is charged false positives that TrackEval forgives. On a synthetic four-row sequence, sa

Description Add a trackers download CLI subcommand that lets users download benchmark datasets from a GCP bucket. Supports downloading full datasets or individual splits, with or without video frames. Datasets: MOT17, SportsMOT, SoccerNet Tracking, DenseTrack. Proposed CLI interface # List what's available trackers download --list # Download full dataset trackers download mot17 # Download specific split, annotations only trackers download mot17 --split train --annotations-only # Custom output path trackers download sportsmot --split val -o ./datasets Tasks Research and document the structure of each dataset (MOT17, SportsMOT, SoccerNet Tracking, DenseTrack): available splits, contents per split (frames, GT annotations, detections), directory layout. Prepare zip files for each dataset/split/variant combination and upload them to a GCP bucket. Create a dataset manifest (mapping dataset + split + variant to GCP URLs and metadata) inside the codebase. Register the download subcommand

Description Generate sample videos for each benchmark dataset used in the Tracker Comparison documentation page. Each video takes ground-truth annotations (bounding boxes + IDs from the dataset's TXT files), draws them on the corresponding frames using the same visual style as the Tracker CLI, and exports a video. Three videos per dataset, four datasets total (12 videos). Upload to GitHub for review; the best one per dataset will later be embedded in the docs. Tasks Identify and select three representative sequences from each of the four datasets (MOT17, SportsMOT, SoccerNet Tracking, DenseTrack). 12 sequences total. Write a script that loads ground-truth MOT TXT annotations and corresponding video frames for a given sequence, using the existing parsers in trackers/io/mot.py. Visualize annotations on each frame: draw bounding boxes and labels (tracker ID + class name if available) with black text at top-left, box and label background colored by tracker ID. Match the Tracker CLI vis