🌟 [CVPR'2025 Highlight] - MITracker

The official implementation for the CVPR 2025 paper [MITracker: Multi-View Integration for Visual Object Tracking].

[Project Page] [Model] [RawResults]

MITracker_demo.mp4

Data Preparation

Please complete this form to request authorization for the non-commercial use of MVTrack. Once submitted, you will receive an email containing the download links to access the 80GB dataset.

Place the tracking datasets and organize the data in the following format:

./data
├── MVTrack/
    ├── ashbin1/
    │   ├── ashbin1-1/
    │   │   ├── img
    │   │   ├── attributes.json     # frame level target attributes
    │   │   ├── groundtruth.txt     # x, y, w, h
    │   │   └── invisible.txt       # fully occlusion or out-of-view
    │   ├── ashbin1-2/
    │   ├── ...
    │   └── BEV/
    │       └── xyz_index.txt       # x: [-4000, 4000] (mm), y: [-4000, 4000] (mm), z: [-50, 2950] (mm), voxel indices 
    ├── ashbin3
    ├── bag1
    ├── basketball5
    ...
    ├── calibs.json                 # camera intrinsics and extrinsics (mm)
    ├── test_split.txt
    ├── train_split.txt
    └── val_split.txt
├── got10k/
    ├── test
    ├── train
    └── val

Below is an example of the annotations provided by the MVTrack dataset.

Environment

Installation

conda create -n mitracker python=3.9
conda activate mitracker
bash install.sh

Set project paths

Run the following command to set paths for this project

python tracking/create_default_local_file.py --workspace_dir . --data_dir /data --save_dir ./output

After running this command, you can also modify paths by editing these two files

lib/train/admin/local.py  # paths about training
lib/test/evaluation/local.py  # paths about testing

Training

Download pre-trained DINOv2 ViT-Base weights and put it under $PROJECT_ROOT$/pretrained_networks.

First stage

Train the view-specific feature extraction module using single-view inputs from MVTrack and got10k datasets.

python tracking/train.py --script mitracker_stage1 --config baseline --save_dir ./output --mode multiple --nproc_per_node 2

Replace --config with the desired model config under experiments/mitracker_stage1.

Second stage

To enable multi-view integration module training, replace experiments/mitracker/*.yaml PRETRAIN_PTH in the yaml configuration file with the path to your pretrained checkpoint, such as './output/checkpoints/train/mitracker_stage1/baseline/MITrackerStage1_ep0050.pth.tar'.

python tracking/train.py --script mitracker --config baseline --save_dir ./output --mode multiple --nproc_per_node 2

Evaluation

Test

MVTrack, GMTD or other off-line evaluated benchmarks (modify --dataset correspondingly)

• mvtrack_sot_test for single-view test, mvtrack_mot_test for multi-view test

python tracking/test.py mitracker_stage1 baseline --dataset mvtrack_sot_test --runid 50 --threads 8 --num_gpus 2 --restart 0
python tracking/test.py mitracker baseline --dataset mvtrack_mot_test --runid 40 --threads 8 --num_gpus 2 --restart 0
python tracking/analysis_results.py # need to modify tracker configs and names

Test FLOPs, and Speed

Note: The speeds reported in our paper were tested on a single NVIDIA A100.

python tracking/profile_model.py --script mitracker --config baseline

Acknowledgments

• Thanks to the ODTrack and TrackTacular libraries for enabling quick implementation of our ideas.

Citation

If any parts of our paper and code help your research, please consider citing us and giving a star to our repository.

@article{xu2025mitracker,
  title={MITracker: Multi-View Integration for Visual Object Tracking},
  author={Xu, Mengjie and Zhu, Yitao and Jiang, Haotian and Li, Jiaming and Shen, Zhenrong and Wang, Sheng and Huang, Haolin and Wang, Xinyu and Yang, Qing and Zhang, Han and others},
  journal={arXiv preprint arXiv:2502.20111},
  year={2025}
}