Related Experiment Video
Updated: Jul 9, 2025
06:37
Author Spotlight: Addressing Technical and Subjective Challenges in Measuring Classroom Attention
Published on: December 15, 2023
3.8K
SiamPKHT: Hyperspectral Siamese Tracking Based on Pyramid Shuffle Attention and Knowledge Distillation
Kun Qian1, Shiqing Wang1, Shoujin Zhang1
1School of Artifical Intelligence and Computer Science, Jiangnan University, Wuxi 214122, China.
Sensors (Basel, Switzerland)
|December 9, 2023
Summary
This study introduces SiamPKHT, a new hyperspectral object tracking (HOT) algorithm. SiamPKHT effectively utilizes spectral features for improved object tracking performance, addressing limitations in current Siamese trackers.
Area of Science:
- Computer Vision
- Remote Sensing
- Machine Learning
Background:
- Hyperspectral images offer rich spectral and spatial data beneficial for object tracking.
- Existing Siamese trackers struggle to fully leverage hyperspectral features due to limited data and high dimensionality.
- Training hyperspectral object tracking (HOT) models is challenging due to data scarcity and computational complexity.
Purpose of the Study:
- To propose a novel hyperspectral object tracking (HOT) algorithm, SiamPKHT, that enhances feature extraction and mitigates overfitting.
- To improve the performance of object tracking in hyperspectral imagery by effectively utilizing spectral information.
- To develop a real-time HOT algorithm capable of handling the complexities of high-dimensional hyperspectral data.
Main Methods:
- The SiamPKHT algorithm is developed by enhancing the SiamCAR model with pyramid shuffle attention (PSA) and knowledge distillation (KD).
- PSA module utilizes pyramid convolutions for multiscale feature extraction and shuffle attention for inter-channel and spatial relationship modeling.
- Knowledge distillation (KD) is employed, guided by a pre-trained RGB tracking model, to address overfitting issues in hyperspectral object tracking.
Main Results:
- Experiments on the HOT2022 dataset demonstrate that SiamPKHT outperforms the baseline SiamCAR and other state-of-the-art HOT algorithms.
- SiamPKHT achieves superior performance in hyperspectral object tracking tasks.
- The proposed algorithm meets real-time processing requirements, achieving a speed of 43 frames per second.
Conclusions:
- SiamPKHT effectively addresses the challenges of hyperspectral object tracking by integrating advanced attention mechanisms and knowledge distillation.
- The algorithm demonstrates significant improvements in tracking accuracy and robustness compared to existing methods.
- SiamPKHT offers a promising solution for real-time, high-performance object tracking in hyperspectral imaging applications.
