IEEE Access (Jan 2024)

Sparse Transformer-Based Sequence Generation for Visual Object Tracking

  • Dan Tian,
  • Dong-Xin Liu,
  • Xiao Wang,
  • Ying Hao

DOI
https://doi.org/10.1109/ACCESS.2024.3482468
Journal volume & issue
Vol. 12
pp. 154418 – 154425

Abstract

Read online

In visual object tracking, attention mechanisms can flexibly and efficiently handle complex dependencies and global information, which improves tracking accuracy. However, when dealing with scenarios that contain a large amount of background information or other complex information, its global attention ability can dilute the weight of important information, allocate unnecessary attention to background information, and thus reduce tracking performance. To relieve this problem, this paper proposes a visual object tracking framework based on a sparse transformer. Our tracking framework is a simple encoder-decoder structure that realizes the prediction of the target in an autoregressive manner, eliminating the additional head network and simplifying the tracking architecture. Furthermore, we introduce a Sparse Attention Mechanism (SMA) in the cross-attention layer of the decoder. Unlike traditional attention mechanisms, SMA focuses only on the top K pixel values that are most relevant to the current pixel when calculating attention weights. This allows the model to focus more on key information and improve foreground and background discrimination, resulting in more accurate and robust tracking. We conduct tests on six tracking benchmarks, and the experimental results prove the effectiveness of our method.

Keywords