The official implementation of the NeurIPS 2023 paper: MixFormerV2: Efficient Fully Transformer Tracking.
-
[Sep 22, 2023] MixFormerV2 is accpeted by NeurIPS 2023! ๐
-
[May 31, 2023] We released two versions of the pretrained model, which can be accessed on Google Driver and NJU Box.
-
[May 26, 2023] Code is available now!
MixFormerV2 is a well unified fully transformer tracking model, without any dense convolutional operation and complex score prediction module. We propose four key prediction tokens to capture the correlation between target template and search area.
To further improve efficiency, we present a new distillation paradigm for tracking model, including dense-to-sparse stage and deep-to-shallow stage.
MixFormerV2 works well for different benchmarks and can achieve 70.6% AUC on LaSOT and 57.4% AUC on TNL2k, while keeping 165fps on GPU. To our best knowledge, MixFormerV2-S is the first transformer-based one-stream tracker which achieves real-time running on CPU.
Use the Anaconda
conda create -n mixformer2 python=3.6
conda activate mixformer2
bash install_requirements.sh
Put the tracking datasets in ./data. It should look like:
${MixFormerV2_ROOT}
-- data
-- lasot
|-- airplane
|-- basketball
|-- bear
...
-- got10k
|-- test
|-- train
|-- val
-- coco
|-- annotations
|-- train2017
-- trackingnet
|-- TRAIN_0
|-- TRAIN_1
...
|-- TRAIN_11
|-- TEST
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
Training with multiple GPUs using DDP.
You can follow instructions (in Chinese now) in training.md.
Example scripts can be found in tracking/train_mixformer.sh
.
bash tracking/train_mixformer.sh
- LaSOT/GOT10k-test/TrackingNet/OTB100/UAV123/TNL2k. More details of test settings can be found in
tracking/test_mixformer.sh
.
bash tracking/test_mixformer.sh
- Progressive eliminating version of training.
- Fast version of test forwarding.
Tianhui Song: [email protected]
Yutao Cui: [email protected]
@misc{mixformerv2,
title={MixFormerV2: Efficient Fully Transformer Tracking},
author={Yutao Cui and Tianhui Song and Gangshan Wu and Limin Wang},
year={2023},
eprint={2305.15896},
archivePrefix={arXiv}
}