/FHT

Matlab implemention of FHT

Primary LanguageMATLAB

Matlab implementation of "Exploiting Spatial and Spectral Attention for Fast Hyperspectral Tracking (FHT)" 及改进版 “快速鲁棒高光谱目标跟踪算法 (FRHT)”


Installation

  • Download the repository.

  • Compile mex files running compile.m command: Set opencv_include and opencv_libpath to the correct Opencv paths.

  • Use evaluate_FHT.m script for the visualization of the tracker. Set tracker_path variable to the directory where your source code is and base_path to the directory where you have stored the HSI sequences.

  • To use the FRHT, set the parameters.feature_type in read_default_csr_parameters hgray for the using of one channel grayscale feature, and set parameters.use_PSR True.

Dataset

Download the Whisper dataset at Baidu Drive(key: n616).

Project summary

Compared with visible images, HyperSpectral Images HSI) can take advantage of rich spectral information to better discriminate foreground objects from complicated backgrounds. Despite continuous performance improvement, HSI trackers are always slow in dealing with high-dimensional HSI data. In this paper, we propose a new method named Fast Hyperspectral Tracker (FHT) in the correlation-filter (CF) based framework to track objects fast in hyperspectral videos by exploiting spatial and spectral attention from the HSI data. The spatial attention map adaptively adjusts the filter support to the part of the object suitable for tracking. The spectral attention weights reflect the spectral-wise quality of the learned filters and are used as the feature weighting coefficients in localization. To acheive fast HSI tracking, we extract lightweight hyperspectral features from high-dimensional HSI data. Extensive experiments on the recently released Whisper dataset demonstrated the effectiveness of our work in terms of both accuracy and efficiency.

FHT Comparison to SOTA (PR/SR)

FHT Running results

Comparison with KCF and SAMF. Illustration of false color sequences "card", "coin", "drive", "forest", and "kangaroo" in the Whisper dataset. The colors of bounding box given by ground truth, FHT, SAMF and KCF+CN are blue, green, red and yellow. image

Contact

Any question regarding this work can be addressed to xuqingyu@nudt.edu.cn.