Motion-Guided Dual-Camera Tracker
(MotionDCTrack)

A highly robust tracker for tracking on stereo cameras

ICRA 2025

Motion-Guided Dual-Camera Tracker for Endoscope Tracking and Motion Analysis in a Mechanical Gastric Simulator
Yuelin Zhang, Kim Yan, Chun Ping Lam, Chengyu Fang, Wenxuan Xie, Yufu Qiu, Raymond Shing-Yan Tang, and Shing Shin Cheng
IEEE International Conference on Robotics and Automation (ICRA) 2025

Key Features

• Dual-Camera-Based Tracking: Uses a cross-camera mutual template strategy (CMT) to enhance tracking consistency and adapt to appearance variations.
• Motion-Guided Prediction Head (MMH): Integrates historical motion data using a Mamba-based motion modeling framework, improving tracking under occlusion and light-induced distortions.
• Superior Accuracy: Outperforms state-of-the-art trackers with 42% lower average error and 72% lower maximum error in 3D tracking.
• Endoscopy Motion Analysis: Provides reliable motion metrics to differentiate between expert and novice endoscopists.

Structure overview of the proposed MotionDCTrack

🔥 News

• 2025-04-08: We release the code, weight, and dataset.
• 2025-03-11: We release this repository.
• 2025-01-28: Our paper is accepted by ICRA 2025.

🩻 Datasets

Experiment setup. (a) Self-developed mechanical gastric simulator and installation of dual-camera tracking devices. (b) Flexible gastric endoscope with EMT affixed at its tip to provide the 3D ground truth. (c) Dual camera pairs used in this work and image examples collected by different camera pairs.

We collected and annotated a dataset using a mechanical gastric simulator. For detail about the gastric simulator, please refer to our paper: A Highly Distensible, Cable Loop-Driven Soft Robotic Gastric Simulator for Endoscopy Training at JMRR.

The raw dataset can be downloaded from Google Drive.

To apply the dataset for training, one can directly use our preprocessed dataset, which can be downloaded from Google Drive, together with the annotation json at Google Drive.

The preprocessed dataset and annotation json should be placed under the training_dataset/my_endos folder.

⚙️ Dependencies

Initialize Conda Environment and Clone Repo

git clone https://github.com/PieceZhang/MotionDCTrack
conda create -n motiondctrack python=3.11
conda activate motiondctrack
# If necessary, Replace pytorch-cuda=? with the compatible version of your GPU driver.
conda install pytorch==2.0.0 torchvision==0.15.1 torchaudio==2.0.1 pytorch-cuda=11.8 -c pytorch -c nvidia

Install

pip install -r requirements.txt
python setup.py build_ext --inplace  # Build extensions

🏃 Train

Download pretrained backbones

Download pretrained resnet50.model from here and put them in pretrained_models directory

Train

cd ./experiments/motiondctrack
export PYTHONPATH=$PYTHONPATH:/path/to/siamban
python -m torch.distributed.launch \
    --nproc_per_node=4 \
    --master_port=2333 \
    ../../tools/train.py --cfg config.yaml
# This is a command to train the model with 4 GPUs. Modify the code according to your GPU settings.

🏃‍♂️ Test

The model weight can be downloaded at here. Put the weight in experiments/motiondctrack directory.

Tracking Demo

You can download an example video in test set from here and put them in demo_video directory.

python tools/test.py

For tracking using 6 cameras, using test_multicams.py instead.

python tools/test_multicams.py

Evalutation

cd /experiment/motiondctrack
python ../../tools/eval.py --dir /path/to/your/rawdataset

🔍 Results

Left: Demonstration of the dual-camera tracking comparison. Our tracker not only achieves the most accurate tracking under multiple disturbances but also has the best tracking consistency across the dual cameras. Right: 3D motion trajectory ground truth measured by EMT and estimated 3D motion from different methods.

📎 Citation

If you find our work helpful, please consider citing our paper:

MotionDCTrack: (official link will be updated soon)

@article{zhang2024motion,
  title={Motion-Guided Dual-Camera Tracker for Endoscope Tracking and Motion Analysis in a Mechanical Gastric Simulator},
  author={Zhang, Yuelin and Yan, Kim and Lam, Chun Ping and Fang, Chengyu and Xie, Wenxuan and Qiu, Yufu and Tang, Raymond Shing-Yan and Cheng, Shing Shin},
  journal={arXiv preprint arXiv:2403.05146},
  year={2024}
}

Gastric Simulator:

@article{lam2024highly,
  title={A Highly Distensible, Cable Loop-Driven Soft Robotic Gastric Simulator for Endoscopy Training},
  author={Lam, Chun Ping and Zhang, Yuelin and Yan, Kim and Ding, Qingpeng and Tang, Raymond Shing-Yan and Cheng, Shing Shin},
  journal={Journal of Medical Robotics Research},
  pages={2350005},
  year={2024},
  publisher={World Scientific}
}