Tracking Anything in High Quality

Technical Report:

👋Welcome everyone to contribute and collaborate on HQTrack repository!

Tracking Anything in High Quality (HQTrack) is a framework for high performance video object tracking and segmentation. It mainly consists of a Video Multi-Object Segmenter (VMOS) and a Mask Refiner (MR), can track multiple target objects at the same time and output accurate object masks.

🍺 HQTrack obtains runner-up in the Visual Object Tracking and Segmentaion (VOTS2023) challenge.

📆TODO

• Demo (can be run locally).
• Training codes.
• Interactive WebUI
• Lightweight version for computing-friendly.

📢News

• [2023/8/17] We release the training codes.
• [2023/7/30] We provide a demo code that can be run locally.
• [2023/7/22] We author a technical report for HQTrack.
• [2023/7/3] HQTrack ranks the 2nd place in the VOTS2023 challenge.

🔥Demo

We also provide a demo script, which supports box and point prompts as inputs. This is a pure python script that allows the user to test arbitrary videos.

🐍Pipeline

📑Intallation

• Install the conda environment

conda create -n hqtrack python=3.8
conda activate hqtrack

• Install Pytorch

conda install pytorch==1.9 torchvision cudatoolkit=10.2 -c pytorch

• Install HQ-SAM

cd segment_anything_hq
pip install -e .
pip install opencv-python pycocotools matplotlib onnxruntime onnx

• Install Pytorch-Correlation-extension package

cd packages/Pytorch-Correlation-extension/
python setup.py install

• Install ops_dcnv3

cd HQTrack/networks/encoders/ops_dcnv3
./make.sh

• Install vots toolkit

pip install vot-toolkit

• Install other packages

pip install easydict
pip install lmdb
pip install einops
pip install jpeg4py
pip install 'protobuf~=3.19.0'
conda install setuptools==58.0.4
pip install timm
pip install tb-nightly
pip install tensorboardx
pip install scikit-image
pip install rsa
pip install six
pip install pillow

🚗Run HQTrack

• Model Preparation

Download VMOS model from Google Driver or Baidu Driver and put it under

/path/to/HQTrack/result/default_InternT_MSDeAOTL_V2/YTB_DAV_VIP/ckpt/

Download HQ-SAM_h and put it under

/path/to/HQTrack/segment_anything_hq/pretrained_model/

• Initialize the vots workspace

cd /path/to/VOTS23_workspace
vot initialize tests/multiobject

• Copy our trackers.ini to your vot workspace

cp /path/to/our/trackers.ini /path/to/VOTS23_workspace/trackers.ini

• Modify your path in trackers.ini
• test the tracker and pack the results

bash run.sh

🐬Training

Stage 1

In stage 1, we pre-train VMOS on synthetic video sequences generated from static image datasets. We refer readers to AFB-URR for preparing the pre-train datasets. The Static dataset should be put in

/path/to/HQTrack/datasets/

• Download pre-trained backbone model and put it under

/path/to/HQTrack/pretrain_models/

• Transfer the backbone ckpt to meet VMOS model

python my_tools/transfer_intern_pretrained_model.py

• Set the relevant training args in

/path/to/HQTrack/configs/pre.py

• Start stage 1 pre-training by running:

CUDA_VISIBLE_DEVICES="1" python tools/train.py --amp \
	--exp_name "Static_Pre" \
	--stage "pre" \
	--model "internT_msdeaotl_v2" \
	--gpu_num "1"

Stage 2

In stage 2, video multi-object segmentation datasets are employed for training, e.g., DAVIS and YoutubeVOS.

• Prepare the datasets and put them under

/path/to/HQTrack/datasets/

• Start stage 2 training by running:

CUDA_VISIBLE_DEVICES="1" python tools/train.py --amp \
	--exp_name "default" \
	--stage "ytb_vip_dav_deaot_internT" \
	--model "internT_msdeaotl_v2" \
	--gpu_num "1"

You can include more training datasets such as VIPSeg, BURST, MOTS, and OVIS for better performance.

📖 Citation

If you find HQTrack useful for you, please consider citing 📣

@misc{hqtrack,
      title={Tracking Anything in High Quality}, 
      Author = {Jiawen Zhu and Zhenyu Chen and Zeqi Hao and Shijie Chang and Lu Zhang and Dong Wang and Huchuan Lu and Bin Luo and Jun-Yan He and Jin-Peng Lan and Hanyuan Chen and Chenyang Li},
      Title = {Tracking Anything in High Quality},
      Year = {2023},
      Eprint = {arXiv:2307.13974},
      PrimaryClass={cs.CV}
}

♥️ Acknowledgment

This project is based on DeAOT, HQ-SAM, and SAM. Thanks for these excellent works.

📧Contact

If you have any question, feel free to email [email protected]. ^_^