HSOD-BIT-V2: A New Challenging Benchmarkfor Hyperspectral Salient Object Detection

TL;DR

This work addresses the limitations of RGB-based salient object detection in challenging scenes by leveraging hyperspectral imagery. It introduces HSOD-BIT-V2, the largest HSOD benchmark to date with 500 hyperspectral images, eight natural backgrounds, five challenging attributes, and a focus on small objects and spectral similarity; 406 images are for training and 94 for testing, with 417 particularly challenging samples. The proposed Hyper-HRNet combines Hyperspectral Attention Reconstruction (HAR) and Global Ternary Perception Decoder (GTPD) to preserve spectral information while improving contour localization, through modules like MSSA, ASAM, CMFI, GAFA, and TAW, and is trained with a hybrid loss that unifies spectral reconstruction, saliency, and global guidance. Empirical results show Hyper-HRNet surpassing both RGB- and HSOD-based baselines on HSOD-BIT-V2, HSOD-BIT, and HS-SOD datasets, demonstrating the effectiveness of spectral-aware reconstruction and global-tertiary decoding in challenging hyperspectral scenes. The work establishes a new benchmark and provides a scalable framework for robust HSOD, with potential impact on real-world vision systems requiring accurate object delineation under spectral variability.

Abstract

Salient Object Detection (SOD) is crucial in computer vision, yet RGB-based methods face limitations in challenging scenes, such as small objects and similar color features. Hyperspectral images provide a promising solution for more accurate Hyperspectral Salient Object Detection (HSOD) by abundant spectral information, while HSOD methods are hindered by the lack of extensive and available datasets. In this context, we introduce HSOD-BIT-V2, the largest and most challenging HSOD benchmark dataset to date. Five distinct challenges focusing on small objects and foreground-background similarity are designed to emphasize spectral advantages and real-world complexity. To tackle these challenges, we propose Hyper-HRNet, a high-resolution HSOD network. Hyper-HRNet effectively extracts, integrates, and preserves effective spectral information while reducing dimensionality by capturing the self-similar spectral features. Additionally, it conveys fine details and precisely locates object contours by incorporating comprehensive global information and detailed object saliency representations. Experimental analysis demonstrates that Hyper-HRNet outperforms existing models, especially in challenging scenarios.

Figures (7)

• Figure 1: (a) Exemplary challenging scenarios from HSOD-BIT-V2, where objects are hard to identify in pseudo-color images but exhibit salience in spectral curves. (b) Challenge attributes of HSOD-BIT paper45 and HSOD-BIT-V2, highlighting their capability to represent real-world challenges.
• Figure 2: Examples of pseudo-color images and corresponding ground truth from HSOD-BIT-V2.
• Figure 3: Diagram of HSOD-BIT-V2 statistics.
• Figure 4: The overall architecture of the proposed Hyper-HRNet is shown in the top part of the figure. The bottom part illustrates the detailed elucidation of composition within HPSAB and the blocks within GTPD.
• Figure 5: Qualitative results on HSOD-BIT-V2 and HSOD-BIT datasets. Hyper-HRNet has best detection performance.