LiteTracker: Leveraging Temporal Causality for Accurate Low-latency Tissue Tracking

Official code repository for LiteTracker: Leveraging Temporal Causality for Accurate Low-latency Tissue Tracking; published at MICCAI 2025.

📑 arXiv

We propose LiteTracker, a low-latency method for tissue tracking in endoscopic video streams. LiteTracker builds on a state-of-the-art long-term point tracking method, and introduces a set of training-free runtime optimizations. These optimizations enable online, frame-by-frame tracking by leveraging a temporal memory buffer for efficient feature reuse and utilizing prior motion for accurate track initialization. LiteTracker demonstrates significant runtime improvements being around 7x faster than its predecessor and 2x than the state-of-the-art. Beyond its primary focus on efficiency, LiteTracker delivers high-accuracy tracking and occlusion prediction, performing competitively on both the STIR and SuPer datasets.

Getting Started

1. Install the required packages using pip (tested only with Ubuntu 20.04 and 22.04 with Python 3.10):

pip install -r requirements.txt

2. Download the pre-trained weights or train your own CoTracker3 Online model via the official repository. In our experiments, we used the scaled weights from the official repository.

3. For evaluation, download the STIR Challenge 2024 and Super datasets:

   • STIR Challenge 2024 and dataset
   • SuPer Framework website and dataset

Inference and Evaluation

Demo and runtime analysis

The demo script runs LiteTracker on a video in a stream-line fashion, produces a video with tracking results, and prints the runtime statistics.

python demo.py -w <path/to/weights.pth> -v assets/stir-5-seq-01.mp4 -s 20 -q 0

STIR Challenge 2024

STIR evaluation scripts are based on the official repository. Minimal modifications are made to accommodate within our framework.

bash ./launch_eval_stir.sh <path/to/STIRDataset> <path/to/weights.pth>

SuPer

python ./src/eval/super/eval_super.py -d <path/to/SuPerDataset> -w <path/to/weights.pth>

Unpublished Results on Non-surgical Datasets

LiteTracker also performs competitively on natural scenes benchmarks. Latency values are computed on a single RTX3090, as 95th percentile of the measurements, tracking 1024 points.

Method	DAVIS			RGB-S			Kinetics			RoboTAP			Dynamic Replica			Latency (ms) ↓
	AJ↑	δ_avg^vis↑	OA↑	AJ↑	δ_avg^vis↑	OA↑	AJ↑	δ_avg^vis↑	OA↑	AJ↑	δ_avg^vis↑	OA↑	δ_avg^occ↑	δ_avg^vis↑	Surv.↑
CoTracker3 (Online)	63.8	76.3	90.2	71.7	83.6	91.1	55.8	68.5	88.3	66.4	78.8	90.8	40.1	73.3	94.4	200.98
TrackOn (48)	65.0	78.0	90.8	71.4	85.2	91.7	53.9	67.3	87.8	63.5	76.4	89.4	-	73.6	-	74.80
LiteTracker	62.0	74.6	88.4	71.0	81.7	86.7	54.4	66.9	85.7	63.6	76.7	87.5	38.5	71.5	93.9	29.67

Acknowledgements

Special thanks to the authors of CoTracker3, MFT, STIR Challenge, and SuPer Framework that made this work possible.

Please cite our work if you use LiteTracker in your research:

@inproceedings{karaoglu2025litetracker,
  title={LiteTracker: Leveraging Temporal Causality for Accurate Low-Latency Tissue Tracking},
  author={Karaoglu, Mert Asim and Ji, Wenbo and Abbas, Ahmed and Navab, Nassir and Busam, Benjamin and Ladikos, Alexander},
  booktitle={International Conference on Medical Image Computing and Computer-Assisted Intervention},
  pages={308--317},
  year={2025},
  organization={Springer}
}