This repository contains the implementation used for:

• LITE: A Paradigm Shift in Multi-Object Tracking with Efficient ReID Feature Integration, ICONIP 2024
• Practical Evaluation Framework for Real-Time Multi-Object Tracking: Achieving Optimal and Realistic Performance, IEEE Access 2025

Authors: Jumabek Alikhanov, Dilshod Obidov, Mirsaid Abdurasulov, Hakil Kim

Paper PDFs are included in docs/LITE.pdf and docs/Practical_Evaluation_Framework_for_Real-Time_Multi-Object_Tracking_Achieving_Optimal_and_Realistic_Performance.pdf.

LITE (Lightweight Integrated Tracking-Feature Extraction) extracts appearance features directly from intermediate YOLO detector features instead of running a separate ReID network for every detection. The repository supports standard trackers and their LITE variants:

• DeepSORT and LITE-DeepSORT
• StrongSORT and LITE-StrongSORT
• Deep OC-SORT and LITE-Deep OC-SORT
• BoT-SORT and LITE-BoT-SORT
• SORT, ByteTrack, and OC-SORT baselines

The paper results use YOLOv8m, input resolution 1280, person class only, confidence threshold 0.25, and layer14 as the default LITE appearance feature layer unless otherwise stated.

Small FPS differences are expected across GPUs, CUDA versions, storage speed, and display/visualization settings. HOTA should be compared using the same TrackEval version, dataset split, confidence threshold, input size, and tracker configuration.

The experiments were developed for Python 3.10 with CUDA-enabled PyTorch. A clean setup is recommended:

git clone https://github.com/Jumabek/LITE.git
cd LITE

python3.10 -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip setuptools wheel

# Choose the PyTorch wheel that matches your CUDA driver.
# This example uses CUDA 12.1.
pip install torch==2.2.2 torchvision==0.17.2 --index-url https://download.pytorch.org/whl/cu121
pip install -r requirements.txt

# Make local LITE, ultralytics, and yolo_tracking modules importable.
export PYTHONPATH="$PWD:$PWD/yolo_tracking:$PYTHONPATH"

If you use a different CUDA version, install the matching torch and torchvision wheels first, then run pip install -r requirements.txt.

The repository includes local copies of the modified ultralytics and yolo_tracking code. For MOT metrics, clone TrackEval next to the project code:

git clone https://github.com/JonathonLuiten/TrackEval.git

The evaluation commands below assume TrackEval/ is in the repository root. If it is elsewhere, replace TrackEval/scripts/run_mot_challenge.py with your local path.

Download the prepared datasets with:

bash scripts/download_datasets_from_gdrive.sh

After extraction, the expected structure is:

datasets/
  MOT17/
    train/MOT17-02-FRCNN/img1/...
    train/MOT17-02-FRCNN/gt/gt.txt
    test/...
  MOT20/
    train/MOT20-01/img1/...
    train/MOT20-01/gt/gt.txt
    test/...
  PersonPath22/test/...
  VIRAT-S/train/...
  KITTI/train/...

You can also download MOT17/MOT20 from the official MOTChallenge site and arrange the folders in the same layout.

Download and extract ReID checkpoints with:

bash scripts/prepare_checkpoints.sh

The expected checkpoint layout is:

checkpoints/
  FastReID/
    bagtricks_S50.yml
    Base-bagtricks.yml
    DukeMTMC_BoT-S50.pth
    deepsort/
      ckpt.t7
      original_ckpt.t7

YOLO weights such as yolov8m.pt are loaded from the repository root or downloaded automatically by Ultralytics when available.

Use the included video to verify the environment before running full benchmarks:

python demo.py --source demo/VIRAT_S_010204_07_000942_000989.mp4

This opens an OpenCV display window. Press q to stop. On a headless server, run the benchmark commands below instead.

The main wrapper is scripts/run_experiment.sh. It writes MOT-format predictions to:

results/paper/<DATASET>-<SPLIT>/<TRACKER>__input_<SIZE>__conf_<CONF>__model_<YOLO>/data/<SEQUENCE>.txt

Run a single paper-setting experiment:

bash scripts/run_experiment.sh \
  -d MOT20 \
  -s train \
  -t LITEStrongSORT \
  -m yolov8m \
  -r 1280 \
  -c 0.25 \
  -o results/paper

Run all MOT17/MOT20 trackers used in the table:

for dataset in MOT17 MOT20; do
  for tracker in DeepSORT StrongSORT DeepOCSORT BoTSORT LITEDeepSORT LITEStrongSORT LITEDeepOCSORT LITEBoTSORT; do
    bash scripts/run_experiment.sh \
      -d "$dataset" \
      -s train \
      -t "$tracker" \
      -m yolov8m \
      -r 1280 \
      -c 0.25 \
      -o results/paper
  done
done

Equivalent direct command for LITE-DeepSORT:

python run.py \
  --dataset MOT20 \
  --split train \
  --tracker_name LITEDeepSORT \
  --input_resolution 1280 \
  --min_confidence 0.25 \
  --appearance_feature_layer layer14 \
  --yolo_model yolov8m \
  --dir_save results/paper/MOT20-train/LITEDeepSORT__input_1280__conf_0.25__model_yolov8m

For BoT-SORT and Deep OC-SORT variants, the wrapper calls yolo_tracking/tracking/run.py with the corresponding --tracking-method and optional --appearance-feature-layer layer14.

Evaluate one dataset split with TrackEval:

python TrackEval/scripts/run_mot_challenge.py \
  --BENCHMARK MOT20 \
  --SPLIT_TO_EVAL train \
  --TRACKERS_FOLDER results/paper/MOT20-train \
  --GT_FOLDER datasets/MOT20/train \
  --GT_LOC_FORMAT "{gt_folder}/{seq}/gt/gt.txt" \
  --METRICS HOTA CLEAR Identity VACE \
  --USE_PARALLEL True \
  --NUM_PARALLEL_CORES 8 \
  --OUTPUT_SUMMARY True \
  --OUTPUT_DETAILED True \
  --PLOT_CURVES True

For MOT17, replace MOT20 with MOT17 and use results/paper/MOT17-train plus datasets/MOT17/train.

TrackEval writes summary files inside each tracker result directory. Use the HOTA, MOTA, and IDF1 columns from those summaries when comparing with the paper table.

The benchmark commands print per-sequence speed and write fps.csv for the run.py based trackers. For a controlled FPS comparison, keep the same GPU, CUDA version, image size, confidence threshold, batch behavior, and visualization setting across all trackers.

Example single-sequence FPS run:

bash scripts/fps.sh -d MOT20 -s train -q MOT20-01

FPS outputs are saved below:

results/<DATASET>-FPS/<SEQUENCE>/<TRACKER>__input_1280__conf_.25/fps.csv

Run the ReID evaluator on one sequence:

python reid.py \
  --dataset MOT20 \
  --seq_name MOT20-01 \
  --split train \
  --tracker LITE \
  --appearance_feature_layer layer14 \
  --output_path reid_results/MOT20-01 \
  --save

Run all LITE layers:

bash scripts/run_reid_all_layers.sh

The evaluator saves ROC and similarity-distribution plots under the selected --output_path.

Download extra demo videos:

bash demo/download_solutions_demo_videos.sh

Run object counting and heatmap demos:

python solutions.py --source videos/shortened_enterance.mp4 --solution object_counter
python solutions.py --source videos/shortened_enterance.mp4 --solution heatmap

Run parking management:

python solutions.py --source videos/parking.mp4 --solution parking_management

Outputs are written to demo_output_videos/.

• ModuleNotFoundError for local modules: run export PYTHONPATH="$PWD:$PWD/yolo_tracking:$PYTHONPATH" from the repository root.
• CUDA or faiss-gpu installation errors: install a PyTorch/CUDA combination supported by your driver, or use a CUDA-enabled conda environment.
• Missing TrackEval: clone TrackEval into the repository root or update the evaluation command path.
• Missing dataset sequence: verify the exact folder names in datasets/<DATASET>/<split>/; MOT17 uses names such as MOT17-02-FRCNN.
• GUI/OpenCV errors on a server: use non-demo benchmark commands or configure a virtual display.

@misc{alikhanov2024liteparadigmshiftmultiobject,
  title={LITE: A Paradigm Shift in Multi-Object Tracking with Efficient ReID Feature Integration},
  author={Jumabek Alikhanov and Dilshod Obidov and Hakil Kim},
  year={2024},
  eprint={2409.04187},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2409.04187}
}

@ARTICLE{10883969,
  author={Alikhanov, Jumabek and Obidov, Dilshod and Abdurasulov, Mirsaid and Kim, Hakil},
  journal={IEEE Access},
  title={Practical Evaluation Framework for Real-Time Multi-Object Tracking: Achieving Optimal and Realistic Performance},
  year={2025},
  volume={13},
  pages={34768-34788},
  doi={10.1109/ACCESS.2025.3541177}
}