SymTrack

Beyond Detection: A Structure-Aware Framework for Scene Text Tracking

Official implementation of SymTrack, accepted by ICML 2026.

News

• [2026-05-01] Code, models, and benchmarks will be released.
• [2026-05-01] SymTrack is accepted by ICML 2026.

Overview

Scene text tracking aims to localize a specified text instance across video frames. Unlike generic visual object tracking, scene text tracking is uniquely challenging because text instances are highly sensitive to geometric deformation, semantic ambiguity, and fine-grained structural changes.

SymTrack is a detection-free scene text tracking framework designed specifically for robust video text tracking. It avoids repeated per-frame detection and instead tracks a target text instance from an initial annotation. SymTrack integrates text-specific structural priors into a unified tracking framework and improves robustness under perspective shifts, dense distractors, and low-quality visual conditions.

Motivation

Modern single object trackers perform well on generic visual targets, but they are not designed for scene text. Scene text has several properties that make direct application of generic trackers unreliable:

1. Perspective-induced distortion.
   Text is usually printed or displayed on planar surfaces. Camera motion and viewpoint changes can cause strong geometric deformation and feature misalignment.

2. High visual ambiguity.
   Different text instances often share similar colors, fonts, strokes, and layouts. Generic trackers can easily drift to nearby distractors.

3. Fine-grained structural sensitivity.
   Small localization errors may change the perceived textual content. Text tracking therefore requires higher structural precision than common object tracking.

SymTrack is built around these observations and introduces a structure-aware tracking paradigm for scene text.

Contributions

• We provide systematic analysis of STT, identifying core challenges including severe distortions from perspective shifts, high visual ambiguity across instances, and fine-grained structural sensitivity.
• We propose SymTrack, a unified architecture equipped with PTR, CEC and AIE, which respectively alleviate structural imbalance, semantic bias, and motion limitation problems of existing trackers.
• Considering the lack of dedicated STT benchmarks, we build upon three datasets from VTS and ensure high-quality annotations. On these benchmarks, our proposed SymTrack sets the new state-of-the-art performance.

Method

Benchmark Results

Type	Method	Venue	ArTVideoSOT AUC (%)	ArTVideoSOT PNorm (%)	ArTVideoSOT P (%)	DSTextSOT AUC (%)	DSTextSOT PNorm (%)	DSTextSOT P (%)	BOVTextSOT AUC (%)	BOVTextSOT PNorm (%)	BOVTextSOT P (%)
Vision-only	SiamRPN++	CVPR2019	56.40	67.30	71.90	44.40	54.40	63.20	58.70	71.50	65.50
Vision-only	STARK	ICCV2021	70.37	83.48	86.84	57.59	68.63	78.01	61.92	75.16	76.33
Vision-only	OSTrack256	ECCV2022	64.86	77.95	81.99	52.50	63.49	70.83	58.67	73.04	74.03
Vision-only	OSTrack384	ECCV2022	64.80	77.82	82.05	54.83	66.51	74.44	59.18	72.68	73.30
Vision-only	AiATrack	ECCV2022	66.41	77.96	81.77	57.92	68.12	79.02	64.16	75.07	73.42
Vision-only	SeqTrackL384	CVPR2023	64.35	76.46	80.92	54.63	65.81	74.19	60.42	76.18	76.70
Vision-only	ARTrack256	CVPR2023	64.85	78.81	79.53	48.53	56.12	65.20	62.75	72.28	73.01
Vision-only	GRM256	CVPR2023	68.22	79.84	83.30	53.05	63.87	71.04	59.59	72.12	72.82
Vision-only	GRM384	CVPR2023	68.47	80.65	83.64	55.51	66.07	74.63	59.13	71.02	71.66
Vision-only	ROMTrack	ICCV2023	70.62	83.32	87.13	56.82	68.79	75.61	62.82	73.74	74.90
Vision-only	ODTrack	AAAI2024	69.81	83.54	86.68	62.71	75.84	84.26	64.74	77.74	78.45
Ours	SymTrack	ICML2026	77.74	91.29	95.88	70.66	83.61	91.83	77.06	90.05	90.18
V-L	DUTrack256	CVPR2025	68.73	82.46	86.87	60.57	72.77	81.31	65.09	78.98	79.04
V-L	DUTrack384	CVPR2025	72.09	85.97	89.36	63.63	76.72	85.00	65.08	79.41	79.30
VTS	TransVTSpotter	NeurIPS2021	8.84	78.11	38.07	4.93	75.21	67.80	-	-	-
VTS	TransDETR	IJCV2024	9.18	78.75	43.31	5.08	76.09	69.79	-	-	-

Visualization

Model

Expected model files after preparation:

${PROJECT_ROOT}
|-- output
|   |-- checkpoints
|   |   |-- train
|   |   |   |-- symtrack
|   |   |   |   |-- baseline_text_scalear
|   |   |   |   |   |-- SymTrack_ep0300.pth.tar
|-- my_internvit_clean
|-- pretrained_networks
|   |-- mae_pretrain_vit_base.pth

Installation

1. Clone this repository

git clone https://github.com/EdisonYCM/SymTrack.git
cd SymTrack

2. Create environment

conda create -n symtrack python=3.9 -y
conda activate symtrack

3. Install dependencies

pip install -U pip
pip install -r requirements.txt

Data Preparation

Download the SymTrack benchmark from ModelScope:

The expected dataset structure is:

${PROJECT_ROOT}
|-- data
|   |-- ArTVideo_SOT_Train
|   |   |-- train
|   |   |   |-- list.txt
|   |   |   |-- sequence_1
|   |   |   |   |-- groundtruth.txt
|   |   |   |   |-- 00000001.jpg
|   |   |   |   |-- ...
|   |   |-- val
|   |   |   |-- list.txt
|   |   |   |-- ...
|   |
|   |-- ArTVideo_SOT_Test
|   |   |-- list.txt
|   |   |-- sequence_1
|   |   |   |-- groundtruth.txt
|   |   |   |-- 00000001.jpg
|   |   |   |-- ...
|   |
|   |-- DSText_SOT_Train
|   |-- DSText_SOT_Test
|   |-- BOVText_SOT_Train
|   |-- BOVText_SOT_Test

Each sequence folder should contain:

sequence_name
|-- groundtruth.txt
|-- 00000001.jpg
|-- 00000002.jpg
|-- ...

The annotation file groundtruth.txt follows the standard SOT format:

x,y,w,h

Download Checkpoints

Option 1: HuggingFace

pip install -U huggingface_hub
hf download EdisonYCM/SymTrack --local-dir ./hf_assets

Option 2: Google Drive

Download the model files from:

https://drive.google.com/drive/folders/1HmathXplBBh0kLqo8NQVy1scQYKeh96D?usp=drive_link

Then organize them as:

${PROJECT_ROOT}
|-- output/checkpoints/train/symtrack/baseline_text_scalear
|-- my_internvit_clean
|-- pretrained_networks

Set Project Paths

Run:

python tracking/create_default_local_file.py \
  --workspace_dir . \
  --data_dir ./data \
  --save_dir ./output

This command generates local path configuration files for training and testing.

Then edit the following files if needed:

lib/train/admin/local.py
lib/test/evaluation/local.py

Evaluation

python tracking/test.py symtrack baseline_text_scalear \
  --dataset_name XXX \
  --runid 300 \
  --threads 8 \
  --num_gpus 1

The results will be saved under:

output/test/tracking_results/

Analyze Results

python tracking/analysis_results.py

Training

python tracking/train.py \
  --script symtrack \
  --config baseline_text_scalear \
  --save_dir ./output \
  --mode multiple \
  --nproc_per_node 4 \

Test FLOPs and Speed

python tracking/profile_model.py --config baseline_text_scalear

---

Citation

If this project is useful for your research, please consider citing:

@inproceedings{yu2026symtrack,
  title={Beyond Detection: A Structure-Aware Framework for Scene Text Tracking},
  author={Yu, Chenmin and Yu, Liu and Wu, Daiqing and Li, Gengluo and Chen, Zeyu and Zhou, Yu},
  booktitle={Proceedings of the 43rd International Conference on Machine Learning},
  year={2026}
}

Acknowledgements

This repository is built upon excellent open-source tracking frameworks and codebases. We sincerely thank the authors of:

• STARK
• OSTrack
• ODTrack
• PyTracking

We also thank the maintainers of the video text spotting datasets used to construct the scene text tracking benchmarks.

License

Licensed under a Creative Commons Attribution-NonCommercial 4.0 International for Non-commercial use only. Any commercial use should get formal permission first.