TrackLab is an easy-to-use modular framework for multi-object pose/segmentation/bbox tracking that supports many tracking datasets and evaluation metrics.
- [2024.02.05] Public release
- Public release of the codebase
- Add support for more datasets (DanceTrack, MOTChallenge, SportsMOT, ...)
- Add many more SOTA tracking methods and object detectors
- Improve documentation and add more tutorials
The TrackLab library is in its early stages, and we're eager to evolve it into a robust, mature tracking framework that can benefit the wider community. If you're interested in contributing, feel free to open a pull-request or reach out to us!
Welcome to this official repository of TrackLab, a modular framework for multi-object tracking. TrackLab is designed for research purposes and supports many types of detectors (bounding boxes, pose, segmentation), datasets and evaluation metrics. Every component of TrackLab, such as detector, tracker, re-identifier, etc, is configurable via standard yaml files (Hydra config framework) TrackLab is designed to be easily extended to support new methods.
TrackLab is composed of multiple modules:
- A detector (YOLOv8, ...)
- A re-identification model (BPBReID, ...)
- A tracker (DeepSORT, StrongSORT, OC-SORT, ...)
Here's what makes TrackLab different from other existing tracking frameworks:
- Fully modular framework to quickly integrate any detection/reid/tracking method or develop your own
- It allows supervised training of the ReID model on the tracking training set
- It provides a fully configurable visualization tool with the possibility to display any dev/debug information
- It supports online and offline tracking methods (compared to MMTracking, AlphaPose, LightTrack and other libs who only support online tracking)
- It supports many tracking-related tasks:
- multi-object (bbox) tracking
- multi-person pose tracking
- multi-person pose estimation
- person re-identification
You can find the documentation at https://trackinglaboratory.github.io/tracklab/ or in the docs/ folder. After installing, you can run make html
inside this folder
to get an html version of the documentation.
Installation guide1
git clone https://github.com/TrackingLaboratory/tracklab.git
cd tracklab
conda create -n tracklab pip python=3.10 pytorch==1.13.1 torchvision==0.14.1 pytorch-cuda=11.7 -c pytorch -c nvidia -y
conda activate tracklab
You might need to change your torch installation depending on your hardware. Please check on Pytorch website to find the right version for you.
Get into your repo and install the requirements with :
pip install -e .
mim install mmcv==2.0.1
You might need to redo this if you update the repository, and some dependencies changed.
- Get the SoccerNet Tracking dataset here, rename the root folder as "SoccerNetMOT" and put it under the global dataset directory (specified under the
data_dir
config as explained below). Otherwise, you can modify thedataset_path
config in soccernet_mot.yaml with your custom SoccerNet dataset directory. - Download the pretrained model weights here and put the "pretrained_models" directory under the main project directory (i.e. "/path/to/tracklab/pretrained_models").
You will need to set up some variables before running the code :
- In configs/config.yaml :
data_dir
: the directory where you will store the different datasets (must be an absolute path !)- All the parameters under the "Machine configuration" header
- In the corresponding modules (tracklab/configs/modules/.../....yaml) :
- The
batch_size
- You might want to change the model hyperparameters
- The
To launch TrackLab with the default configuration defined in configs/config.yaml, simply run:
tracklab
This command will create a directory called outputs
which will have a ${experiment_name}/yyyy-mm-dd/hh-mm-ss/
structure.
All the output files (logs, models, visualization, ...) from a run will be put inside this directory.
If you want to override some configuration parameters, e.g. to use another detection module or dataset, you can do so by modifying the corresponding parameters directly in the .yaml files under configs/.
All parameters are also configurable from the command-line, e.g. : (more info on Hydra's override grammar here)
tracklab 'data_dir=${project_dir}/data' 'model_dir=${project_dir}/models' modules/reid=bpbreid pipeline=[bbox_detector,reid,track]
${project_dir}
is a variable that is configured to be the root of the project you're running the code in. When using
it in a command, make sure to use single quotes (') as they would otherwise be seen as
environment variables.
To find all the (many) configuration options you have, use :
tracklab --help
The first section contains the configuration groups, while the second section shows all the possible options you can modify.
TODO Describe TrackLab + Hydra configuration system
Here is an overview of the important TrackLab classes:
- TrackingDataset: Abstract class to be instantiated when adding a new dataset. The
TrackingDataset
contains oneTrackingSet
for each split of the dataset (train, val, test, etc).- Example: SoccerNetMOT. The SoccerNet Tracking dataset.
- TrackingSet: A tracking set contains three Pandas dataframes:
video_metadatas
: contains one row of information per video (e.g. fps, width, height, etc).image_metadatas
: contains one row of information per image (e.g. frame_id, video_id, etc).detections_gt
: contains one row of information per ground truth detection (e.g. frame_id, video_id, bbox_ltwh, track_id, etc).
- TrackerState: Core class that contains all the information about the current state of the tracker. All modules in the tracking pipeline update the tracker_state sequentially. The tracker_state contains one key dataframe:
detections_pred
: contains one row of information per predicted detection (e.g. frame_id, video_id, bbox_ltwh, track_id, reid embedding, etc).
- TrackingEngine: This class is responsible for executing the entire tracking pipeline on the dataset. It loops over all videos of the dataset and calls all modules defined in the pipeline sequentially. The exact execution order (e.g. online/offline/...) is defined by the TrackingEngine subclass.
- Example: OfflineTrackingEngine. The offline tracking engine performs tracking one module after another to speed up inference by leveraging large batch sizes and maximum GPU utilization. For instance, YoloV8 is first applied on an entire video by batching multiple images, then the re-identification model is applied on all detections in the video, etc.
- Pipeline: Define the order in which modules are executed by the TrackingEngine. If a tracker_state is loaded from disk, modules that should not be executed again must be removed.
- Example: [bbox_detector, reid, track]
- VideoLevelModule: Abstract class to be instantiated when adding a new tracking module that operates on all frames simultaneously. Can be used to implement offline tracking strategies, tracklet level voting mechanisms, etc.
- Example: VotingTrackletJerseyNumber. To perform majority voting within each tracklet and compute a consistent tracklet level attribute (an attribute can be, for instance, the result of a detection level classification task).
- ImageLevelModule: Abstract class to be instantiated when adding a new tracking module that operates on a single frame. Can be used to implement online tracking strategies, pose/segmentation/bbox detectors, etc.
- Example 1: YOLOv8. To perform object detection on each image with YOLOv8. Creates a new row (i.e. detection) within
detections_pred
. - Example 2: StrongSORT. To perform online tracking with StrongSORT. Creates a new "track_id" column for each detection within
detections_pred
.
- Example 1: YOLOv8. To perform object detection on each image with YOLOv8. Creates a new row (i.e. detection) within
- DetectionLevelModule: Abstract class to be instantiated when adding a new tracking module that operates on a single detection. Can be used to implement pose estimation for top-down strategies, re-identification, attributes recognition, etc.
- Callback: Implement this class to add a callback that is triggered at a specific point during the tracking process, e.g. when dataset/video/module processing starts/ends.
- Example: VisualizationEngine. Implements "on_video_loop_end" to save each video tracking results as a .mp4 or a list of .jpg.
- Evaluator: Implement this class to add a new evaluation metric, such as MOTA, HOTA, or any other (non-tracking related) metrics.
- Example: TrackEvalEvaluator. Evaluate the performance of a tracker using the official TrackEval library.
Here is an overview of what happen when you run TrackLab:
tracklab/main.py is the main entry point and receives the complete Hydra's configuration as input.
tracklab/main.py is usually called via the following command through the root main.py file: python main.py
.
Within tracklab/main.py, all modules are first instantiated.
Then training any tracking module (e.g. the re-identification model) on the tracking training set is supported by calling the "train" method of the corresponding module.
Tracking is then performed on the validation or test set (depending on the configuration) via the TrackingEngine.run() function.
For each video in the evaluated set, the TrackingEngine calls the "run" method of each module (e.g. detector, re-identifier, tracker, ...) sequentially.
The TrackingEngine is responsible for batching the input data (e.g. images, detections, ...) before calling the "run" method of each module with the correct input data.
After a module has been called with a batch of input data, the TrackingEngine then updates the TrackerState object with the module outputs.
At the end of the tracking process, the TrackerState object contains the tracking results of each video.
Visualizations (e.g. .mp4 results videos) are generated during the TrackingEngine.run() call, after a video has been tracked and before the next video is processed.
Finally, evaluation is performed via the evaluator.run() function once the TrackingEngine.run() call is completed, i.e. after all videos have been processed.
When developing a new module, it is often useful to dump the tracker state to disk to save computation time and avoid running the other modules several times. Here is how to do it:
- First, save the tracker state by using the corresponding configuration in the config.yaml file:
defaults:
- state: save
# ...
state:
save_file: "states/${experiment_name}.pklz" # 'null' to disable saving. This is the save path for the tracker_state object that contains all modules outputs (bboxes, reid embeddings, jersey numbers, roles, teams, etc)
load_file: null
- Run Tracklab. The tracker state will be saved in the experiment folder as a .pklz file.
- Then modify the load_file key in "config.yaml" to specify the path to the tracker state file that has just been created (
load_file: "..."
config). - In config.yaml, remove from the pipeline all modules that should not be executed again. For instance, if you want to use the detections and reid embeddings from the saved tracker state, remove the "bbox_detector" and "reid" modules from the pipeline. Use
pipeline: []
if no module should be run again.
defaults:
- state: save
# ...
pipeline:
- track
# ...
state:
save_file: null # 'null' to disable saving. This is the save path for the tracker_state object that contains all modules outputs (bboxes, reid embeddings, jersey numbers, roles, teams, etc)
load_file: "path/to/tracker_state.pklz"
- Run Tracklab again.
If you use this repository for your research or wish to refer to our contributions, please use the following BibTeX entries:
@misc{Joos2024Tracklab,
title = {{TrackLab}},
author = {Joos, Victor and Somers, Vladimir and Standaert, Baptiste},
journal = {GitHub repository},
year = {2024},
howpublished = {\url{https://github.com/TrackingLaboratory/tracklab}}
}
SoccerNet Game State Reconstruction:
@inproceedings{Somers2024SoccerNetGameState,
title = {{SoccerNet} Game State Reconstruction: End-to-End Athlete Tracking and Identification on a Minimap},
author = {Somers, Vladimir and Joos, Victor and Giancola, Silvio and Cioppa, Anthony and Ghasemzadeh, Seyed Abolfazl and Magera, Floriane and Standaert, Baptiste and Mansourian, Amir Mohammad and Zhou, Xin and Kasaei, Shohreh and Ghanem, Bernard and Alahi, Alexandre and Van Droogenbroeck, Marc and De Vleeschouwer, Christophe},
booktitle = cvsports,
month = Jun,
year = {2024},
address = city-seattle,
}
@article{bpbreid,
archivePrefix = {arXiv},
arxivId = {2211.03679},
author = {Somers, Vladimir and {De Vleeschouwer}, Christophe and Alahi, Alexandre},
doi = {10.48550/arxiv.2211.03679},
eprint = {2211.03679},
file = {:Users/vladimirsomers/Library/CloudStorage/OneDrive-SportradarAG/PhD/Scientific Literature/Mendeley/Somers, De Vleeschouwer, Alahi_2023_Body Part-Based Representation Learning for Occluded Person Re-Identification.pdf:pdf},
isbn = {2211.03679v1},
journal = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV23)},
month = {nov},
title = {{Body Part-Based Representation Learning for Occluded Person Re-Identification}},
url = {https://arxiv.org/abs/2211.03679v1 http://arxiv.org/abs/2211.03679},
year = {2023}
}
Footnotes
-
Tested on
conda 22.11.1
,Python 3.10.8
,pip 22.3.1
,g++ 11.3.0
andgcc 11.3.0
↩