BoxMOT Files

BoxMOT v26.0.0

BoxMOT v26.0.0 adds multimodal tracking, optional EdgeTAM mask guidance, typed tracker configuration, and joint motion-model selection. It also improves evaluation reproducibility, replay performance, and tuning diagnostics.

Based on committed changes since v25.0.0, through 9da023fd.

New features and improvements

• EagerMOT sensor fusion: Track objects using calibrated 2D and 3D detections, with optional camera poses. New spatial structures represent 3D boxes, camera calibration, and multimodal outputs. Includes KITTI evaluation, class-specific tuning, and 3D Kalman calibration.
• MAF-HDA mask tracking: Adds a Python tracker combining motion with mask-based correlation-filter appearance. Supports AABB detections with instance masks without requiring ReID weights.
• EdgeTAM integration: Generate detection masks with the standalone segmentor or enable temporal mask-guided association using --edgetam. Guidance supports the nine Python box trackers with AABB geometry, IoU association, and per_class=False.
• Typed configuration and simpler factories: All eleven trackers expose immutable, validated configuration classes with editor completion. Detector, ReID, and tracker factories support simpler construction, while shared KalmanConfig and ReIDConfig group component settings.
• Joint motion selection: tune --select-kf compares compatible Kalman filters and noise profiles using held-out predictions. Add --select-cmc to jointly select camera-motion compensation. Selected profiles are saved for reuse.
• More targeted tuning: Conditional search spaces account for geometry, association metric, and enabled features. New controls expose motion, detection-confidence, and ReID–IoU fusion, with improved trial comparisons and parameter diagnostics.
• Expanded evaluation: Adds KITTI-MOTS mask evaluation, saved 2D detection workflows, 3D tracking metrics, and optional official KITTI AP40 reports. Materialized AABB evaluation supports ordered GTA, GSI, and GBRC postprocessing while preserving raw results.
• Faster repeated workloads: Batched Kalman operations, optimized camera-motion compensation, indexed embedding reads, reusable workers, and optional --cache-inputs reduce repeated processing.
• Reliability fixes: Restores per-sequence evaluation reporting, corrects DeepOC-SORT covariance handling during camera compensation, cleans up orphaned workers, and fixes Python 3.10 typing compatibility. OpenCV is pinned to 4.11.0.86, and its runtime version participates in perception-cache identity.

Breaking changes

This major release includes breaking changes. When upgrading from v25:

• Update direct tracker construction. Pass algorithm settings through the matching config, such as BotSort(config=BotSortConfig(...)). The factory also accepts algorithm keywords: create_tracker("botsort", match_thresh=0.8).
• Group component settings. Move flat Kalman options into kalman and kalman.noise; configure tracker-owned appearance through reid=ReIDConfig(...) or an injected appearance encoder.
• Replace track_buffer. ByteTrack and BoT-SORT now use max_lost_frames, counted directly in tracker updates without FPS scaling.
• Review removed tracker features. SAM2MOT has been removed; standalone SAM segmentation remains available. OccluBoost's online GTA controls are removed, and its AABB/OBB thresholds and lifetime settings are consolidated.
• Review centroid thresholds. centroid_max_distance now determines geometric eligibility across matching and recovery; overlap thresholds no longer restrict centroid matching.
• Update commands and imports. Replace boxmot time-variant with boxmot materialize --time-variant. Internal tracker and motion modules have moved; prefer public imports from boxmot.
• Rebuild native libraries. ByteTrack, BoT-SORT, and OccluBoost use constructor ABI v3. Their shared update/result ABI remains v2.
• Install optional dependencies explicitly. Use boxmot install --extra ...; first-use automatic installation has been removed. Published Docker images are now CPU-only, while GPU build targets remain available.

EdgeTAM currently requires its full PyTorch checkpoint. The registered export-edgetam command and TFLite bundle runtime remain unfinished.