:unicorn: This is a Generic Object Tracking Project

April 26, 2026 · View on GitHub

🔥 Paper Accepted at ICLR 2026!

GOT-Edit: Geometry-Aware Generic Object Tracking via Online Model Editing

📄 Paper and Project Page

Human perception for object tracking in a 2D video stream arises from the implicit use of prior visual geometry 🛰️ and semantic reasoning 👁️. GOT-Edit aligns with this principle by enabling trackers to infer 3D geometry from 2D streaming inputs for visual tracking.

The core of this work is cross-modality online model editing. This mechanism performs online constrained model updates to incorporate geometric information adaptively while preserving semantic discrimination for online adaptation under streaming 2D inputs. This paradigm is generalizable across diverse scenarios and environments 🌐. We hope these advances chart a path toward reliability, safety, and social responsibility in vision systems.

Raw Results and Models

The raw results and models are available for download here

DatasetModelNPrSucPrOP50OP75
NfS-30ToMP-5084.0066.8680.5884.3653.50
PiVOT-L86.6668.2284.5386.0555.45
GOT-Edit87.4771.1286.6489.3059.83
LaSOTToMP-5077.9867.5772.2479.7965.06
PiVOT-L84.6873.3782.0985.6475.18
GOT-Edit85.0875.3183.1786.1377.52
AVisTToMP-5066.6651.6147.7459.4738.88
PiVOT-L81.2062.1865.5573.2555.46
GOT-Edit82.5064.4568.2674.3559.68
OTB-100ToMP-5085.9870.0790.8387.8357.79
PiVOT-L88.4671.2094.5889.3555.73
GOT-Edit91.4774.9697.4293.0263.22

Suc: Success Rate AUC
Pr: Precision AUC
NPr: Normalise Precision AUC

Prerequisites

The codebase is built based on PyTracking.

Familiarity with the PyTracking codebase will help in understanding the structure of this project.

Installation

Clone the GIT repository.

git clone https://github.com/chenshihfang/GOT.git

Ensure that CUDA 11.7 is installed. We use PyTorch version 2.0.0.

Install dependencies

sudo apt-get install libturbojpeg

Install Environment

Run the installation script to install all the dependencies. You need to provide the conda install path and the name for the created conda environment

bash install_gotedit.sh /your_anaconda3_path/ gotedit
conda activate gotedit

Install geometry dependencies

  1. Change to the Depth-Anything-3 directory:

    cd /yourpath/GOT/pytracking/ltr/Depth-Anything-3/  
python -m pip install -e . --no-deps

    Other geometry backbones can be installed in a similar way.

  2. Change back to the pytracking directory:

    cd /home/sfchen94/GOT/pytracking/

Set Up the Dataset Environment

You can follow the setup instructions from PyTracking.

There are two different local.py files located in:

  • ltr/admin
  • pytracking/evaluation

Set Up the Checkpoint Environment

Updating the checkpoint path in ltr/models/backbone/resnet.py is required. This file includes function calls for both the semantic and geometry backbones.

Evaluate the Tracking Performance Based on Datasets

python evaluate_GOT_Edit_results.py

For the GOT-10K and TrackingNet results, please refer to the public leaderboards on the official evaluation websites for both challenges under the entry named “Edit” or “GOT-Edit.” The NfS results follow the evaluation protocol described here.

Run the Tracker

CUDA_VISIBLE_DEVICES=0 python pytracking/run_experiment.py myexperiments_gotedit GOT_Edit --debug 0 --threads 1

Training script

Change directory to GOT/pytracking/:

cd GOT/pytracking/
CUDA_VISIBLE_DEVICES=0,1,2,3 python ltr/run_training_dsA.py tomp GOT-Edit_DA3_378

Consider citing “GOT-Edit” if this project impresses you

@inproceedings{gotedit2026iclr,
title     = {{GOT}-Edit: Geometry-Aware Generic Object Tracking via Online Model Editing},
author    = {Shih-Fang Chen and Jun-Cheng Chen and I-Hong Jhuo and Yen-Yu Lin},
booktitle = {Proc. Int. Conf. Learn. Represent. (ICLR)},
year      = {2026}
}

:fire: GOT-JEPA has been accepted at TCSVT 2026! 👇

A learning framework that enables dynamic model adaptation in adverse environments and fine-grained occlusion perception.

The current repository includes important code for GOT-JEPA; search this repository for tompnet_JEPA.py, tompnet_JEPAp.py, and tompnet_PT.py, which correspond to GOT-JEPA pretraining, GOT-JEPA fine-tuning, and the OccuSolver variant, respectively; also search for heads.py for related implementation details and actors/tracking.py for loss details; usage instructions will be continuously updated.

Raw Results

The raw results can be downloaded from here.

:fire: PiVOT has been accepted at TMM 2025! 👇

PiVOT proposes a prompt generation network with the pre-trained foundation model CLIP to automatically generate and refine visual prompts, enabling the transfer of foundation model knowledge for tracking.

Raw Results

The raw results can be downloaded from here.

Please visit here for usage details.

Acknowledgement

This codebase is implemented on PyTracking libraries.

Citing

If you find this repository useful, please consider giving a star :star: and a citation

@inproceedings{gotedit2026iclr,
title     = {{GOT}-Edit: Geometry-Aware Generic Object Tracking via Online Model Editing},
author    = {Shih-Fang Chen and Jun-Cheng Chen and I-Hong Jhuo and Yen-Yu Lin},
booktitle = {Proc. Int. Conf. Learn. Represent. (ICLR)},
year      = {2026}
}

@ARTICLE{TCSVT_GOT_JEPA,
title={{GOT}-JEPA: Generic Object Tracking with Model Adaptation and Occlusion Handling using Joint-Embedding Predictive Architecture}, 
author={Chen, Shih-Fang and Chen, Jun-Cheng and Jhuo, I-Hong and Lin, Yen-Yu},
journal={IEEE Transactions on Circuits and Systems for Video Technology}, 
year={2026},
doi={10.1109/TCSVT.2026.3675005}
}

@ARTICLE{TMM_PiVOT,
title={Improving Visual Object Tracking Through Visual Prompting}, 
author={Chen, Shih-Fang and Chen, Jun-Cheng and Jhuo, I-Hong and Lin, Yen-Yu},
journal={IEEE Transactions on Multimedia}, 
year={2025},
volume={27},
pages={2682-2694},
doi={10.1109/TMM.2025.3535323}}

Contact:

mail: csf.cs09@nycu.edu.tw or shihfang1207@gmail.com