Related Experiment Video
Updated: Oct 10, 2025
12:39
A Methodology for Capturing Joint Visual Attention Using Mobile Eye-Trackers
Published on: January 18, 2020
7.8K
Learning Geometry Information of Target for Visual Object Tracking with Siamese Networks
Hang Chen1, Weiguo Zhang1, Danghui Yan1
1Automation College, Northwestern Polytechnical University, Xi'an 710072, China.
Sensors (Basel, Switzerland)
|December 10, 2021
Summary
This study introduces a Siamese deformable cross-correlation network for visual tracking. It effectively models target geometry and appearance variations, significantly improving tracking performance.
Area of Science:
- Computer Vision
- Machine Learning
Background:
- Siamese networks are prevalent in visual tracking, often using cross-correlation.
- Standard cross-correlation struggles with background noise, missing foreground information, and target geometry.
Purpose of the Study:
- To propose a novel Siamese deformable cross-correlation network for enhanced visual tracking.
- To address limitations of traditional cross-correlation in handling target deformation and geometric information.
Main Methods:
- Developed a Siamese deformable cross-correlation network learning an offset field end-to-end.
- Integrated an online classification sub-network to model target appearance variations.
- Evaluated on OTB2015, VOT2018, VOT2019, and UAV123 benchmarks.
Main Results:
- The proposed method effectively models target geometric structure through adaptive sampling.
- The online classification sub-network enhances tracker robustness against appearance changes.
- Achieved state-of-the-art performance across multiple challenging visual tracking datasets.
Conclusions:
- Siamese deformable cross-correlation networks offer a significant advancement in visual tracking.
- The approach successfully mitigates background noise and foreground information loss.
- Demonstrated superior performance and robustness in complex tracking scenarios.
