Two-stream Beats One-stream: Asymmetric Siamese Network for Efficient Visual Tracking [AAAI25]

Official implementation of AsymTrack, including models and training&testing codes.

Models & Raw Results (Google Driver) Models & Raw Results (Baidu Driver: asym)

🔥 This work proposes AsymTrack, a new asymmetric siamese framework for efficient visual tracking.

News

[Mar 1, 2025]

• We release codes, models and raw results.
  Thanks for your star.

[Dec 10, 2024]

• AsymTrack is accepted to AAAI2025.

Introduction

• 🔥 A new asymmetric siamese framework for efficient visual tracking, beats the current prevailing one-stream pipeline.

• An efficient template modulation (ETM) mechanism for relation modeling, paired with an object perception enhancement (OPE) module.

• AsymTrack series, tailored for resource-constrained platforms (GPU & CPU & edge AI devices), leads in both accuracy and speed.

• We envision the AsymTrack family becoming a dependable visual tracking solution for real-world deployment, bridging the gap between academic research and industrial applications.

Performance

AsymTrack outperforms state-of-the-art efficient trackers and achieves superior running speed across diverse computing platforms. This approach broadens the possibilities for real-time visual tracking on resource-constrained platforms, offering a viable alternative to one-stream architectures.

Usage

Installation

Create and activate a conda environment:

conda create -n asymtrack python=3.9
conda activate asymtrack

Install the required packages:

bash install_asymtrack.sh

Data Preparation

Put the training datasets in ./data/. It should look like:

${AsymTrack_ROOT}
-- data
  -- lasot
      |-- airplane
      |-- basketball
      |-- bear
      ...
  -- got10k
      |-- test
      |-- train
      |-- val
  -- coco
      |-- annotations
      |-- images
  -- trackingnet
      |-- TRAIN_0
      |-- TRAIN_1
      ...
      |-- TRAIN_11
      |-- TEST

Path Setting

Run the following command to set paths:

cd <PATH_of_ViPT>
python tracking/create_default_local_file.py --workspace_dir . --data_dir ./data --save_dir ./output

You can also modify paths by these two files:

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

Training

Dowmload the pretrained backbone model and put it under ./pretrained_model/.

python -m torch.distributed.launch --nproc_per_node 2 lib/train/run_training.py --script AsymTrack --config base --save_dir .

(Optionally) Debugging training with a single GPU

python lib/train/run_training.py --script AsymTrack --config base --save_dir .

Testing and evaluation on benchmarks

GOT10K

python tracking/test.py AsymTrack base --dataset got10k_test --threads 2 --num_gpus 2
python lib/test/utils/transform_got10k.py --tracker_name AsymTrack --cfg_name base

LaSOT

python tracking/test.py AsymTrack base --dataset lasot --threads 2 --num_gpus 2
python tracking/analysis_results.py --dataset_name lasot --tracker_name AsymTrack --tracker_version base

TrackingNet

python tracking/test.py AsymTrack base --dataset trackingnet --threads 2 --num_gpus 2
python lib/test/utils/transform_trackingnet.py --tracker_name AsymTrack --cfg_name base 

Speed Testing

python tracking/profile_model_asymtrack.py --script AsymTrack --config tiny

Bixtex

If you find AsymTrack is helpful for your research, please consider citing:

  @inproceedings{AsymTrack,
  title={Two-stream Beats One-stream: Asymmetric Siamese Network for Efficient Visual Tracking},
  author={Jiawen, Zhu and Huayi, Tang and Xin, Chen and Xinying, Wang and Dong, Wang and Huchuan, Lu},
  booktitle={AAAI},
  year={2025}}

Acknowledgment

• This repo is based on HiT which is an excellent work.
• We thank for the PyTracking library, which helps us to quickly implement our ideas.