Causal HyperPrompter: A Framework for Unbiased Hyperspectral Camouflaged Object Tracking.

Area of Science:

• Computer Vision
• Machine Learning
• Remote Sensing

Background:

• Hyperspectral camouflaged object tracking faces challenges due to high object-replica similarity.
• Existing trackers suffer from bias and confounding effects from RGB-based fine-tuning and inadequate token-type embeddings.

Purpose of the Study:

• To develop an unbiased tracking framework for hyperspectral camouflaged objects.
• To mitigate bias inherited from RGB models and improve semantic links between template and search tokens.

Main Methods:

• Introduced a structural causal model and counterfactual intervention to disentangle causal factors and eliminate confounding variables.
• Developed a novel token-type embedding module integrating local spectral angle modeling.
• Created the large-scale BihoT-130k hyperspectral camouflaged object detection and tracking dataset.

Main Results:

• The proposed Causal HyperPrompter framework effectively mitigates bias in hyperspectral camouflaged object tracking.
• The new token-type embedding enhances semantic links, improving object localization sensitivity.
• Extensive experiments demonstrate significant improvements across multiple large-scale datasets.

Conclusions:

• Causal HyperPrompter offers a robust solution for unbiased hyperspectral camouflaged object tracking.
• The developed dataset facilitates further research and development in this challenging domain.