Release v13.0.7

Fix multiprocessing CUDA in val.py: #2022

Release v13.0.6 - CLI

• Extended supported Python version range
• torch and torchvision GPU install on Linux and Windows and CPU on Mac
• Always used CPU torch and torchvision in CI, leading to speedups of 20/30 seconds per job

Release v13.0.5 - CLI

• Updates dependency indexes and dependency-to-index mapping
• Changed build system.
• Introduces a new attribute “ch” with a hardcoded value to detectors
• Simplifies the sed command in multiple CI steps to replace the index value.

Release v13.0.4 - CLI

Fix bug: #2008

Release v13.0.3 - CLI

Fix bug: #2008

Release v13.0.2 - CLI

• tracking classes default to 0 for evaluation
• tracking classes default to all for tracking

Release v13.0.1 - CLI

Fix CLI related bugs. Now this works:

pip install boxmot

and then

boxmot track \
  --yolo-model yolov8n.pt \
  --reid-model osnet_x0_25_msmt17.pt \
  --imgsz 640 \
  --tracking-method boosttrack \
  --source /content/test_5.avi \
  --save \
  --show-trajectories \
  --conf 0.1 \
  --show

Release v13.0.0 - CLI

BoxMOT now provides a unified CLI boxmot with the following subcommands:

Usage: boxmot COMMAND [ARGS]...

Commands:
  track                  Run tracking only
  generate-dets-embs     Generate detections and embeddings
  generate-mot-results   Generate trackeval compatible results based on pre-generated detections and embeddings
  eval                   Runs generate-dets-embs and generate-mot-results, finishing with an official trackeval evaluation
  tune                   Tune tracker hyperparameters based on selected detections and embeddings

Release v12.0.10

• val.py now uses ProcessPoolExecutor for true parallelism (no GIL contention). Benchmark run on Apple M3 Max (macOS 15.4):

python3 tracking/val.py \
    --classes 0 \
    --yolo-model yolox_x_ablation.pth \
    --reid-model lmbn_n_duke.pt \
    --tracking-method x \
    --verbose \
    --source ./tracking/val_utils/data/MOT17-50/train/

Tracking Method	Previous Runtime	Current Runtime
StrongSORT	~15 minutes	37 seconds
BoostTrack	1 m 46 s	6 seconds

• Export packages have been moved out of individual .py files and consolidated into pyproject.toml.
  • Simplifies dependency management
  • Ensures consistent versioning across all export workflows

• Import cleanups by @ahmadmughees

• Update OpenVino ReID export

• Ultralytics package now points to the original one, not the custom one anymore

Release v12.0.9

• Unified camera motion compensation utilization cross all trackers
• Camera motion compensation ablation study performed :
  https://github.com/mikel-brostrom/boxmot/wiki/BoostTrack-warp-mode-experiments
• Based on ablation study new argument are used:
  • warp mode: cv2.MOTION_TRANSLATION
  • scale factor: 0.15

New ecc config

Tracker	Status	HOTA↑	MOTA↑	IDF1↑	FPS
boosttrack	✅	69.253	75.914	83.206	25
botsort	✅	68.885	78.222	81.344	46
strongsort	✅	68.05	76.185	80.763	17
bytetrack	✅	67.68	78.039	79.157	1265
deepocsort	✅	67.509	75.83	79.976	12
ocsort	✅	66.441	74.548	77.899	1483

Old ecc config

Tracker	Status	HOTA↑	MOTA↑	IDF1↑	FPS
boosttrack	✅	69.015	75.957	82.538	25
botsort	✅	68.367	78.321	80.7	46
strongsort	✅	68.038	76.142	80.731	17
bytetrack	✅	67.68	78.039	79.157	1265
deepocsort	✅	67.509	75.83	79.976	12
ocsort	✅	66.441	74.548	77.899	1483