RELO: Reinforcement Learning to Localize for Visual Object Tracking

中文说明

This repository contains the public ICML release for RELO. Public training and inference use the relo_warmup and relo experiment families.

Paper: arXiv:2605.07379

Installation

Create and activate a conda environment:

conda create -n relo python=3.10 -y
conda activate relo

Install the Python dependencies:

bash install.sh

Data and Local Paths

Generate local path files from the repository root:

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

Place datasets under ./data using the paths created in lib/train/admin/local.py and lib/test/evaluation/local.py. Adjust those files if your dataset layout differs.

Checkpoints

The public checkpoints and required Fast-iTPN encoder weights are hosted at xche32/RELO on Hugging Face. Download all released assets with:

python tools/download_checkpoints.py --variant all

To download only one model size, replace all with t256, b256, or l256. See MODEL_ZOO.md for asset paths.

Training

Training uses torchrun. The public YAML files encode the paper schedule: 90 warmup epochs with 100000 samples per epoch, then 90 RELO epochs with 2500 samples per epoch and LR drop at epoch 72.

Warmup training example:

torchrun --nproc_per_node=8 lib/train/run_training.py \
  --script relo_warmup \
  --config relo_warmup_b256 \
  --save_dir .

RELO training example:

torchrun --nproc_per_node=8 lib/train/run_training.py \
  --script relo \
  --config relo_b256 \
  --save_dir .

For T256 or L256 training, replace relo_warmup_b256 and relo_b256 with the matching public config names: relo_warmup_t256, relo_t256, relo_warmup_l256, or relo_l256. RELO training expects the matching warmup checkpoint path from the YAML to exist under checkpoints/train/relo_warmup/..., and warmup training expects the matching Fast-iTPN encoder weight under pretrained/itpn/....

Evaluation

Run RELO-B256 on LaSOT:

python tracking/test.py relo relo_b256 --dataset_name lasot --threads 8 --num_gpus 1

Analyze the results:

python tracking/analysis_results.py relo relo_b256 --dataset_name lasot

Supported dataset_name values include lasot, lasot_extension_subset, trackingnet, got10k_test, tnl2k, nfs, and uav. Dataset-specific evaluation settings such as window factors and template update behavior are encoded in the public YAML configs.

Contact

• Xin Chen (email: xche32@cityu.edu.hk)

Citation

If this work proves helpful to your research, we kindly ask that you consider citing our paper.

@inproceedings{relo,
  title={{RELO}: Reinforcement Learning to Localize for Visual Object Tracking},
  author={Xin Chen and Chuanyu Sun and Jiao Xu and Houwen Peng and Dong Wang and Huchuan Lu and Kede Ma},
  booktitle={International Conference on Machine Learning (ICML)},
  year={2026}
}