Related Experiment Video
Updated: Jul 3, 2026
09:44
Methods to Test Visual Attention Online
Published on: February 19, 2015
12.1K
Siamese anchor-free object tracking with multiscale spatial attentions
Jianming Zhang1,2, Benben Huang3,4, Zi Ye3,4
1School of Computer and Communication Engineering, Changsha University of Science and Technology, Changsha, 410114, China. jmzhang@csust.edu.cn.
Scientific Reports
|November 26, 2021
Summary
This study introduces an anchor-free Siamese object tracking algorithm that enhances spatial information capture using multiscale spatial attentions. The novel approach improves tracking robustness and accuracy without relying on predefined anchors.
Area of Science:
- Computer Vision
- Machine Learning
- Artificial Intelligence
Background:
- Siamese network-based object trackers demonstrate high performance.
- Anchor-based methods improve target prediction but struggle with spatial information and robustness due to pre-defined anchors.
Purpose of the Study:
- To propose a Siamese-based anchor-free object tracking algorithm.
- To enhance spatial information capture and improve tracking robustness.
Main Methods:
- Utilized ResNet-50 as a backbone for multiscale feature generation.
- Introduced a spatial attention extraction (SAE) block to capture spatial information.
- Developed an anchor-free classification and regression subnetwork for direct target localization.
Main Results:
- The proposed tracker effectively captures spatial information using multiscale spatial attentions.
- The anchor-free approach eliminates limitations of pre-defined anchors, enhancing robustness.
- Experimental results on OTB100, UAV123, VOT2016, and GOT-10k benchmarks validate the tracker's effectiveness.
Conclusions:
- The developed Siamese-based anchor-free object tracker with multiscale spatial attentions achieves superior performance.
- The method offers a robust and accurate alternative to existing anchor-based tracking algorithms.
