[CVPR 2023] ARKitTrack: A New Diverse Dataset for Tracking Using Mobile RGB-D Data

Project Page | arXiv

This is a PyTorch implementation of the paper ARKitTrack: A New Diverse Dataset for Tracking Using Mobile RGB-D Data. Code will be released here.

Haojie Zhao^*1, Junsong Chen^*1, Lijun Wang¹, Huchuan Lu^1,2 (* indicates equal contributions)

¹Dalian University of Technology, China, ²Peng Cheng Laboratory, China

Contact at: [email protected], [email protected]

---

News

• Code for VOS is coming soon ...
• [2023/06/08] Release Train sets v1.
• [2023/05/09] Release Test sets v1.
• [2023/04/20] Release code for VOT.

---

Dataset

• VOT_test_set: [BaiduNetdisk], VOS_test_set: [BaiduNetdisk]
• All: [BaiduNetdisk], [OneDrive(Available before 2023.7.15)]

---

1. Installation

# 1. Clone this repo
git clone https://github.com/lawrence-cj/ARKitTrack.git
cd ARKitTrack

# 2. Create conda env
conda env create -f art_env.yml
conda activate art

# 3. Install mmcv-full, mmdet, mmdet3d for the BEV pooling, which is from bevfusion.
pip install openmim
mim install mmcv-full==1.4.0
mim install mmdet==2.20.0
python setup.py develop  # mmdet3d

2. 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 and lib/test/evaluation/local.py.

3. Evaluation

Download our trained models from Google Drive and uncompress them to output/checkpoints/.

Change the corresponding dataset paths in lib/test/evaluation/local.py.

Run the following command to test on different datasets.

python tracking/test.py --tracker art --param vitb_384_mae_ce_32x4_ep300 --dataset depthtrack --threads 2 --num_gpus 2

• --config vitb_384_mae_ce_32x4_ep300 is used for cdtb and depthtrack.
• --config vitb_384_mae_ce_32x4_ep300_art is used for arkittrack.
• --debug 1 for visualization.
• --dataset: [depthtrack, cdtb, arkit].

The raw results are stored in Google Drive.

4. Training

Download the pre-trained weights from Google Drive and uncompress it to pretrained_models/.

Change the corresponding dataset paths in lib/train/admin/local.py.

Run the following command to train for vot.

python tracking/train.py --script art --config vitb_384_mae_ce_32x4_ep300 --save_dir ./output --mode multiple --nproc_per_node 2

• --config vitb_384_mae_ce_32x4_ep300: train with depthtrack, test on cdtb and depthtrack.
• --config vitb_384_mae_ce_32x4_ep300_art: train with arkittrack, test on arkittrack.
• You can modify the config yaml files for your own datasets.

---

Acknowledgments

Thanks for the OSTrack and BEVFusion projects, which help us to quickly implement our ideas.

Citation

@InProceedings{Zhao_2023_CVPR,
    author    = {Zhao, Haojie and Chen, Junsong and Wang, Lijun and Lu, Huchuan},
    title     = {ARKitTrack: A New Diverse Dataset for Tracking Using Mobile RGB-D Data},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2023},
    pages     = {5126-5135}
}

License

This project is under the MIT license. See LICENSE for details.