MixFormer

The official implementation of the CVPR 2022 paper MixFormer: End-to-End Tracking with Iterative Mixed Attention

[Models and Raw results] (Google Driver) [Models and Raw results] (Baidu Driver: hmuv)

News

[Jan 01, 2024]

• Extended version has been accepted to T-PAMI 2024 (https://ieeexplore.ieee.org/document/10380715), we would appreciate if you can CITE our new trackers (i.e., MixViT, MixViT (ConvMAE)) in your tracking works.

[Feb 10, 2023]

• 🔥🔥🔥 Code and models for MixViT and MixViT-ConvMAE are available now ! Thank Tianhui Song for helping us clean up the code.

[Feb 8, 2023]

• Extended version has been available at https://arxiv.org/abs/2302.02814. In particular, the extented MixViT-L(ConvMAE) achieves AUC score of 73.3% on LaSOT. Besides, we design a new TrackMAE pre-training method for tracking. Code and models will be updated soon.

[Oct 26, 2022]

• MixFormerL (based on MixViT-L) rank 1/41 on VOT2022-STb public dataset.
• The VOT2022-RGBD and VOT2022-D winners of MixForRGBD and MixForD, implemented by Lai Simiao, are constructed upon our MixFormer.
• The VOT2022-STs winner of MS-AOT employs MixFormer as a part of the tracker. The VOT2022-STb winner of APMT_MR employs the SPM proposed in MixFormer to select dynamic templates.

[Mar 29, 2022]

• Our paper is selected for an oral presentation.

[Mar 21, 2022]

• MixFormer is accepted to CVPR2022.
• We release Code, models and raw results.

Highlights

✨ New transformer tracking framework

MixFormer is composed of a target-search mixed attention (MAM) based backbone and a simple corner head, yielding a compact tracking pipeline without an explicit integration module.

✨ End-to-end, post-processing-free

Mixformer is an end-to-end tracking framework without post-processing.

✨ Strong performance

Tracker	VOT2020 (EAO)	LaSOT (NP)	GOT-10K (AO)	TrackingNet (NP)
MixViT-L (ConvMAE)	0.567	82.8	-	90.3
MixViT-L	0.584	82.2	75.7	90.2
MixCvT	0.555	79.9	70.7	88.9
ToMP101* (CVPR2022)	-	79.2	-	86.4
SBT-large* (CVPR2022)	0.529	-	70.4	-
SwinTrack* (Arxiv2021)	-	78.6	69.4	88.2
Sim-L/14* (Arxiv2022)	-	79.7	69.8	87.4
STARK (ICCV2021)	0.505	77.0	68.8	86.9
KeepTrack (ICCV2021)	-	77.2	-	-
TransT (CVPR2021)	0.495	73.8	67.1	86.7
TrDiMP (CVPR2021)	-	-	67.1	83.3
Siam R-CNN (CVPR2020)	-	72.2	64.9	85.4
TREG (Arxiv2021)	-	74.1	66.8	83.8

Install the environment

Use the Anaconda

conda create -n mixformer python=3.6
conda activate mixformer
bash install_pytorch17.sh

Data Preparation

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

${MixFormer_ROOT}
 -- data
     -- lasot
         |-- airplane
         |-- basketball
         |-- bear
         ...
     -- got10k
         |-- test
         |-- train
         |-- val
     -- coco
         |-- annotations
         |-- train2017
     -- trackingnet
         |-- TRAIN_0
         |-- TRAIN_1
         ...
         |-- TRAIN_11
         |-- TEST

Set project paths

Run the following command to set paths for this project

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

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

Train MixFormer

Training with multiple GPUs using DDP. More details of other training settings can be found at tracking/train_mixformer_[cvt/vit/convmae].sh for different backbone respectively.

# MixFormer with CVT backbone
bash tracking/train_mixformer_cvt.sh

# MixFormer with ViT backbone
bash tracking/train_mixformer_vit.sh

# MixFormer with ConvMAE backbone
bash tracking/train_mixformer_convmae.sh

Test and evaluate MixFormer on benchmarks

• LaSOT/GOT10k-test/TrackingNet/OTB100/UAV123. More details of test settings can be found at tracking/test_mixformer_[cvt/vit/convmae].sh

bash tracking/test_mixformer_cvt.sh
bash tracking/test_mixformer_vit.sh
bash tracking/test_mixformer_convmae.sh

• VOT2020
  Before evaluating "MixFormer+AR" on VOT2020, please install some extra packages following external/AR/README.md. Also, the VOT toolkit is required to evaluate our tracker. To download and instal VOT toolkit, you can follow this tutorial. For convenience, you can use our example workspaces of VOT toolkit under external/vot20/ by setting trackers.ini.

cd external/vot20/<workspace_dir>
vot evaluate --workspace . MixFormerPython
# generating analysis results
vot analysis --workspace . --nocache

Run MixFormer on your own video

bash tracking/run_video_demo.sh

Compute FLOPs/Params and test speed

bash tracking/profile_mixformer.sh

Visualize attention maps

bash tracking/vis_mixformer_attn.sh

Model Zoo and raw results

The trained models and the raw tracking results are provided in the [Models and Raw results] (Google Driver) or [Models and Raw results] (Baidu Driver: hmuv).

Contact

Yutao Cui: cuiyutao@smail.nju.edu.cn

Acknowledgments

• Thanks for PyTracking Library and STARK Library, which helps us to quickly implement our ideas.
• We use the implementation of the CvT from the official repo CvT.

✏️ Citation

If you think this project is helpful, please feel free to leave a star⭐️ and cite our paper:

@inproceedings{cui2022mixformer,
  title={Mixformer: End-to-end tracking with iterative mixed attention},
  author={Cui, Yutao and Jiang, Cheng and Wang, Limin and Wu, Gangshan},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={13608--13618},
  year={2022}
}
@ARTICLE{cui2023mixformer,
      journal={IEEE Transactions on Pattern Analysis and Machine Intelligence}, 
      title={MixFormer: End-to-End Tracking with Iterative Mixed Attention}, 
      author={Yutao Cui and Cheng Jiang and Gangshan Wu and Limin Wang},
      year={2024}
}