The official PyTorch implementation of our ICLR 2026 paper:

FARTrack: Fast Autoregressive Visual Tracking with High Performance

GitHub maintainer: Guijie Wang [ICLR 2026] [arxiv]

• Release training code and data.
• Release evaluation code.
• Release checkpoints.

[Baidu Pan] passwd: arks

Google Drive: FARTrack | FARTrackDistill | FARTrackSparse |

[Raw Result]

We have released the training code for FARTrack and evaluation code into this repository.

Inference speed and tracking performance are two critical evaluation metrics in the field of visual tracking. However, high-performance trackers often suffer from slow processing speeds, making them impractical for deployment on resource-constrained devices. To alleviate this issue, we propose FARTrack, a Fast Auto-Regressive Tracking framework. Since autoregression emphasizes the temporal nature of the trajectory sequence, it can maintain high performance while achieving efficient execution across various devices. FARTrack introduces Task-Specific Self-Distillation and Inter-frame Autoregressive Sparsification, designed from the perspectives of shallow-yet-accurate distillation and redundant-to-essential token optimization, respectively. Task-Specific Self-Distillation achieves model compression by distilling task-specific tokens layer by layer, enhancing the model's inference speed while avoiding suboptimal manual teacher-student layer pairs assignments. Meanwhile, Inter-frame Autoregressive Sparsification sequentially condenses multiple templates, avoiding additional runtime overhead while learning a temporally-global optimal sparsification strategy. FARTrack demonstrates outstanding speed and competitive performance. It delivers an AO of 70.6% on GOT-10k in real-time. Beyond, our fastest model achieves a speed of 343 FPS on the GPU and 121 FPS on the CPU.

The inference speed is evaluated with NVIDIA TiTan Xp, Intel(R) Xeon(R) Gold 6230R CPU @ 3.00GHz, and Ascend 310B.

Use the Anaconda (CUDA 12.2)

conda env create -f FARTrack_env_cuda122.yaml

Run the following command to set paths for this project

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

After running this command, you can also modify paths by editing these two files

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

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

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

Download pre-trained MAE ViT-Tiny weights and put it under $PROJECT_ROOT$/pretrained_models (different pretrained models can also be used, see MAE for more details).

Since sequence-level training requires video input, and the COCO dataset contains only images, traditional training methods were first used to train the model so that it could be fairly compared to other trackers.

python tracking/train.py --script fartrack --config fartrack_tiny_224_full --save_dir ./output --mode multiple --nproc_per_node 4 --use_wandb 0

Replace --config with the desired model config under experiments/fartrack. We use wandb to record detailed training logs, in case you don't want to use wandb, set --use_wandb 0.

This phase uses the same datasets and data augmentations as the frame-level pretraining phase but introduces additional KL divergence loss $\mathcal{L}{\mathrm{KL}}$ and trajectory sequence loss $\mathcal{L}{\mathrm{traj}}=\mathcal{L}{\mathrm{CE}}+\mathcal{L}{\mathrm{SIoU}}$ for each layer. This process generates multiple versions of the distilled model, producing models at different layers from a single distillation.

python tracking/train.py --script fartrack_distill --config fartrack_distill_224_full --save_dir ./output --mode multiple --nproc_per_node 4 --use_wandb 0

To enable sequence-level training, replace 'experience/fartrack_sparse/*.yaml' PRETRAIN_PTH in the yaml configuration file with the path to your pretrained checkpoint, such as './output/fartrack_distill_224_full/checkpoints/train/fartrack_distill/fartrack_distill_224_full/FARTrackDistill_ep0240.pth.tar'.

python tracking/train.py --script fartrack_sparse --config fartrack_sparse_224_full --save_dir ./output --mode multiple --nproc_per_node 4 --use_wandb 0

Change the corresponding values of lib/test/evaluation/local.py to the actual benchmark saving paths

Some testing examples:

• LaSOT or other off-line evaluated benchmarks (modify --dataset correspondingly)

python tracking/test.py fartrack_sparse fartrack_sparse_224_full --dataset lasot --threads 16 --num_gpus 4
python tracking/analysis_results.py # need to modify tracker configs and names

• GOT10K-test

python tracking/test.py fartrack_sparse fartrack_sparse_224_full --dataset got10k_test --threads 16 --num_gpus 4
python lib/test/utils/transform_got10k.py --tracker_name fartrack_sparse --cfg_name fartrack_sparse_224_full

• TrackingNet

python tracking/test.py fartrack_sparse fartrack_sparse_224_full --dataset trackingnet --threads 16 --num_gpus 4
python lib/test/utils/transform_trackingnet.py --tracker_name fartrack_sparse --cfg_name fartrack_sparse_224_full

❤️❤️❤️Our idea is implemented base on the following projects. We really appreciate their excellent open-source works!

• SIoU [related paper]
• OSTrack [related paper]
• PyTracking [related paper]
• ARTrack [related paper]
• ARTrackV2 [related paper]

❤️❤️❤️This project is not for commercial use. For commercial use, please contact the author.

❤️❤️❤️This project is not for commercial use. For commercial use, please contact the author.

❤️❤️❤️This project is not for commercial use. For commercial use, please contact the author.

If any parts of our paper and code help your research, please consider citing us and giving a star to our repository.

@inproceedings{
    wang2026fartrack,
    title={{FART}rack: Fast Autoregressive Visual Tracking with High Performance},
    author={Guijie Wang and Tong Lin and Yifan Bai and Anjia Cao and Shiyi Liang and Wangbo Zhao and Xing Wei},
    booktitle={The Fourteenth International Conference on Learning Representations},
    year={2026},
    url={https://openreview.net/forum?id=lq7Zfr8kAS}
}

If you have any questions or concerns, feel free to open issues.