Object Tracking Network Based on Deformable Attention Mechanism
Kexin Chen (Nanjing University of Posts and Telecommunications),* Baojie Fan (Nanjing University of Posts and Telecommunications), xiaobin Guo (Nanjing University of Posts and Telecommunications)
The 33rd British Machine Vision Conference
Abstract
Recently, many Transformer-based algorithms have emerged in the field of object tracking. Thanks to the full Attention mechanism in the Transformer structure, these tracking algorithms have achieved competitive results, but such model parameters are often bloated compared with CNN-based. In this paper, we focus on the characteristics of the object tracking task, explore novel interaction between template frames and search frames, and propose DeTrack. The modified model uses a combination of an encoder module based on deformable attention mechanism and an encoder module based on self-attention mechanism for feature interaction. The deformable attention-based encoder can precisely track the target location without focusing on all the pixels, which reduces the number of model parameters and effectively improves the model accuracy. We have achieved state-of-the-art performance on LaSOT, TrackingNet, GOT-10K and VOT2020.Citation
@inproceedings{Chen_2022_BMVC,
author = {Kexin Chen and Baojie Fan and xiaobin Guo},
title = {Object Tracking Network Based on Deformable Attention Mechanism},
booktitle = {33rd British Machine Vision Conference 2022, {BMVC} 2022, London, UK, November 21-24, 2022},
publisher = {{BMVA} Press},
year = {2022},
url = {https://bmvc2022.mpi-inf.mpg.de/0469.pdf}
}
Copyright © 2022
The British Machine Vision Conference is organised by