Awesome
π BUSCA for MOT
This is the official repository of Lost and Found: Overcoming Detector Failures in Online Multi-Object Tracking.
We propose BUSCA (Building Unmatched trajectorieS Capitalizing on Attention), a compact plug-and-play module that can be integrated with any online Tracker by Detection (TbD) and enhance it. BUSCA is fully-online and requires no retraining.
<div align="center"> <img src="./assets/teaser.png" alt="Teaser TbD vs TbD+BUSCA"> </div>π Abstract
Multi-object tracking (MOT) endeavors to precisely estimate the positions and identities of multiple objects over time. The prevailing approach, tracking-by-detection (TbD), first detects objects and then links detections, resulting in a simple yet effective method. However, contemporary detectors may occasionally miss some objects in certain frames, causing trackers to cease tracking prematurely. To tackle this issue, we propose BUSCA, meaning 'to search', a versatile framework compatible with any online TbD system, enhancing its ability to persistently track those objects missed by the detector, primarily due to occlusions. Remarkably, this is accomplished without modifying past tracking results or accessing future frames, i.e., in a fully online manner. BUSCA generates proposals based on neighboring tracks, motion, and learned tokens. Utilizing a decision Transformer that integrates multimodal visual and spatiotemporal information, it addresses the object-proposal association as a multi-choice question-answering task. BUSCA is trained independently of the underlying tracker, solely on synthetic data, without requiring fine-tuning. Through BUSCA, we showcase consistent performance enhancements across five different trackers and establish a new state-of-the-art baseline across three different benchmarks.
π vs. SOTA
<div align="center"> <img src="./assets/sota.jpg" alt="Comparison with state of the art"> </div>π₯ Download BUSCA repository
You can download the BUSCA repository with the following command:
git clone https://github.com/lorenzovaquero/BUSCA.git
cd BUSCA
π Download base trackers
The official code of BUSCA supports five different multi-object trackers (StrongSORT, ByteTrack, GHOST, TransCenter, and CenterTrack), but its design allows it to be integrated into any online tracker by detection. You can choose which one(s) you want to download with the following commands:
git submodule update --init trackers/StrongSORT/ # https://github.com/dyhBUPT/StrongSORT
git submodule update --init trackers/ByteTrack/ # https://github.com/ifzhang/ByteTrack
git submodule update --init trackers/GHOST/ # https://github.com/dvl-tum/GHOST
git submodule update --init trackers/TransCenter/ # https://gitlab.inria.fr/robotlearn/TransCenter_official
git submodule update --init trackers/CenterTrack/ # https://github.com/xingyizhou/CenterTrack
π οΈ Environment setup
BUSCA's dependencies can be found on requirements.txt
. You can install them using A) a Docker container or B) directly through pip.
-
A) Docker container
The
Dockerfile
we provide contains all the dependencies for BUSCA and the 5 supported baseline trackers. In order to build and run the container, we provide the scriptsbuild.sh
andrun_docker.sh
:./build.sh # Creates the image named "busca" ./run_docker.sh # Instantiates a container named "busca_container"
Keep in mind that, for building the Docker image, you will need GPU access during build (specifically, for TransCenter's MSDA and DCN). This can be achieved by using
nvidia-container-runtime
. You can find a short guide here and some troubleshooting tips here. If none of this works, you can still proceed with the image creation and later manually install the TransCenter dependencies at runtime, inside the container itself (you will be prompted with a warning when you start the container). -
B) pip install
The dependencies of BUSCA can be installed on the host machine with the following command (using a virtual environment like venv or conda is recommended):
pip3 install -r requirements.txt -f https://download.pytorch.org/whl/cu115 # You may select your specific version of CUDA
Keep in mind that, if you follow the pip approach, you will need to manually install the dependencies for the trackers you may want to use.
πΎ Download BUSCA weights
The trainable components of BUSCA are the appearance feature extractor and the Decision Transformer. If you want to use our pretrained weights, you can find them on the following Google Drive link.
You just need to download them and put model_busca.pth
in the folder models/BUSCA/motsynth/
and put model_feats.pth
in the folder models/feature_extractor/market1501/
(you may need to create the folder structure at the root of the repository). This can be done with the following command:
mkdir -p models/BUSCA/motsynth
gdown 15LB6SPHtDc-4_fLQtRIzU1YWOTF6vNf -O models/BUSCA/motsynth/model_busca.pth # model_busca.pth
mkdir -p models/feature_extractor/market1501
gdown 1ZNU0yNkhMTlLRSOC0PR82SwK1ic9OJ8Y -O models/feature_extractor/market1501/model_feats.pth # model_feats.pth
ποΈ Data preparation
You can download the data for MOT17 and MOT20 from the official website.
Each tracker expects slightly different folder structures and data preprocessing (please, check their respective repositories), but most of our scripts expect that the data will be organized as shown (you can edit the path to the datasets folder in run_docker.sh
):
/beegfs/datasets
βββ MOT17
β βββ test
β βββ train
βββ MOT20
βββ test
βββ train
π Using BUSCA
In order to use BUSCA, you have to 1) download the base tracker, 2) setup the tracker and install its requirements, 3) apply BUSCA to it, and 4) run the experiment. You can choose your favorite tracker, and in this example we will do it using StrongSORT.
1) Download base tracker
You can download StrongSORT from its official repository using our git submodule
.
git submodule update --init trackers/StrongSORT/ # https://github.com/dyhBUPT/StrongSORT
2) Setup base tracker
Each tracker has different setup processes and requirements. Please, check their official repositories for more detailed steps.
In the case of StrongSORT, you first need to install its requirements (this part is not necessary if you are using the Docker image we provide).
Secondly, you have to download its prepared data from Google Drive or Baidu Wangpan (code "sort
"). You can do it with the following command:
mkdir -p trackers/StrongSORT/Dataspace
gdown 1I1Sk6a1i377UhXykqN9jZmbvJLc6izbz -O trackers/StrongSORT/Dataspace/MOT17_ECC_val.json # MOT17_ECC_val.json
gdown 1zzzUROXYXt8NjxO1WUcwSzqD-nn7rPNr -O trackers/StrongSORT/Dataspace/MOT17_val_YOLOX+BoT --folder # MOT17_val_YOLOX+BoT
3) Apply BUSCA
In order to apply BUSCA to a tracker, you simply have to copy and overwrite the files found in the adapters
folder. For our example, you can do it for StrongSORT as follows:
cp -TRv adapters/StrongSORT trackers/StrongSORT
4) Run experiment
You can run the experiments for the different trackers using the scripts found in the scripts
folder:
./scripts/run_strongsort.sh --dataset MOT17 --testset val
The results will be located in the exp
folder. Once computed, the tracking metrics for StrongSORT+BUSCA on MOT17-val should be as follows:
Model | MOTAβ | HOTAβ | IDF1β | IDsβ |
---|---|---|---|---|
StrongSORT | 76.174 | 69.289 | 81.864 | 234 |
StrongSORT+BUSCA | 76.795 | 69.392 | 82.272 | 219 |
π’ Tracking metrics
You can measure the MOTA/HOTA/IDF1 metrics of the different trackers using the official code of TrackEval.
π¦Ύ Training BUSCA
Coming soon! π
π¬ Qualitative test set results
<div align="center"> <img src="./assets/StrongSORT_MOT20-08.gif" alt="StrongSORT+BUSCA on MOT20-08"><img src="./assets/ByteTrack_MOT20-04.gif" alt="ByteTrack+BUSCA on MOT20-04"> <img src="./assets/TransCenter_MOT17-03.gif" alt="TransCenter+BUSCA on MOT17-03"><img src="./assets/CenterTrack_MOT17-11.gif" alt="CenterTrack+BUSCA on MOT17-11"> </div>π Notes
- You can apply BUSCA to other trackers! We would be very happy if you could report your new results!
π Citation
If you find BUSCA useful, please star the project and consider citing us as:
@inproceedings{Vaquero2024BUSCA,
author = {Lorenzo Vaquero and
Yihong Xu and
Xavier Alameda-Pineda and
V{\'{\i}}ctor M. Brea and
Manuel Mucientes},
title = {Lost and Found: Overcoming Detector Failures in Online Multi-Object Tracking},
booktitle = {European Conf. Comput. Vis. ({ECCV})},
series = {Lecture Notes in Computer Science},
volume = {15131},
pages = {448--466},
publisher = {Springer},
year = {2024},
doi = {10.1007/978-3-031-73464-9\_27}
}