GPTrack

📈 LasHeR Result Comparison

The following OPE curves compare GPTrack with representative RGB-T trackers on the LasHeR testing set.

View the original PDF

✨ Highlights

• Geometry-Saliency Cross-Attention (GSCA): Integrates imaging-inspired guidance into cross-modal interaction.
• Multi-Relation Topology Graph (MRTG): Builds spatial, cross-modal, and semantic edges for graph-enhanced RGB-T representation learning.
• LasHeR-ready Training and Evaluation: Provides experiment configs and scripts for RGB-T tracking workflows.

🗂️ Repository Structure

GPTrack/
├── assets/                    # README figures and result curves
├── experiments/gptrack/       # Training and evaluation configs
├── lib/
│   ├── config/gptrack/        # GPTrack configuration
│   ├── models/gptrack/        # Model builder and tracking utilities
│   ├── train/                 # Training actors, datasets, and trainers
│   └── test/                  # Evaluation datasets and tracker wrappers
├── pretrained_models/         # Local pretrained weights
└── tracking/                  # Training, testing, and analysis entry points

⚙️ Installation

conda create -n gptrack python=3.8
conda activate gptrack
bash install.sh

🧭 Path Setup

Initialize local paths for datasets, checkpoints, and results:

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

You can further customize paths in:

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

🗃️ Data Preparation

Place RGB-T datasets under ./data. For LasHeR, the expected layout is:

data/
└── lasher/
    ├── trainingset/
    ├── testingset/
    ├── trainingsetList.txt
    └── testingsetList.txt

🧩 Pretrained Weights

Place pretrained weights under:

pretrained_models/

The default configs expect SOT/ViT initialization weights to be available locally.

🚀 Training

Train GPTrack on LasHeR:

python tracking/train.py \
  --script gptrack \
  --config vitb_256_gptrack_32x1_1e4_lasher_15ep_sot \
  --save_dir ./output/vitb_256_gptrack_32x1_1e4_lasher_15ep_sot \
  --mode multiple \
  --nproc_per_node 4

Available experiment configs are stored in:

experiments/gptrack/

📊 Evaluation

Run tracking on the LasHeR test split:

python tracking/test.py \
  gptrack \
  vitb_256_gptrack_32x1_1e4_lasher_15ep_sot \
  --dataset_name lasher_test \
  --threads 6 \
  --num_gpus 1

Analyze tracking results:

python tracking/analysis_results.py \
  --tracker_name gptrack \
  --tracker_param vitb_256_gptrack_32x1_1e4_lasher_15ep_sot \
  --dataset_name lasher_test

🧱 Core Components

• lib/models/gptrack/gptrack.py: GPTrack model builder and tracking wrapper.
• lib/models/gptrack/vit_gptrack_backbone.py: Main implementation of GSCA and MRTG details within the ViT backbone.
• lib/models/gptrack/utils.py: token conversion utilities and multi-relation edge construction.
• experiments/gptrack/: configuration files for training and evaluation.

🙏 Acknowledgements

The ViT backbone is initialized from the pretrained OSTrack checkpoint. We appreciate the OSTrack project for its valuable open-source resources.

📝 Citation

If GPTrack is useful for your research, please cite the related work and this repository.