Research map / Detection and segmentation / CNN features
HSOD-BIT-V2: A Challenging Benchmark for Hyperspectral Salient Object Detection
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces HSOD-BIT-V2, a 500-hyperspectral-image salient object detection benchmark with five challenge attributes, and Hyper-HRNet, an HRNet-based network with spectral attention reconstruction and a global ternary perception decoder.
Problem
Hyperspectral salient object detection (HSOD) is limited by scarce, small, and low-quality datasets and by spectral redundancy, weak spectral feature capture, and poor contour delineation in existing methods.
Approach
A new large HSOD dataset emphasizing small objects and foreground-background similarity, plus a high-resolution network that reduces spectral dimensionality via CNN-Transformer hybrid spectral attention reconstruction and decodes with global saliency and ternary (object/background/uncertain) contour-aware weights.
Key results
- On HSOD-BIT-V2, Hyper-HRNet reaches MAE 0.028, avgF1 0.589, AUC 0.988 and CC 0.737, outperforming SMN and U2Net.
- On HSOD-BIT, Hyper-HRNet achieves AUC 0.996 and CC 0.916 versus SMN's 0.963 and 0.846.
- On HS-SOD, Hyper-HRNet reaches AUC 0.953 and CC 0.810, improving AUC by 0.037 over SMN.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Hyper-HRNet instead of SMN | HSOD-BIT-V2 · HSOD-BIT-V2 | MAE: 0.039 -> 0.028 | Helped |
| Use Hyper-HRNet instead of U2Net | HSOD-BIT-V2 · HSOD-BIT-V2 | AUC: 0.948 -> 0.988 | Helped |
| Use Hyper-HRNet instead of SMN | HSOD-BIT · HSOD-BIT | AUC: 0.963 -> 0.996 | Helped |
| Use Hyper-HRNet instead of DMSSN | HS-SOD · HS-SOD | MAE: 0.068 -> 0.056 | Helped |
| Add GTPD to the HRNet with HAR (ablation) | HRNet with HAR · HSOD-BIT-V2 | AUC: 0.883 -> 0.988 | Helped |
| Add HAR to the HRNet with GTPD (ablation) | HRNet with GTPD · HSOD-BIT-V2 | AUC: 0.890 -> 0.988 | Helped |
| Use interpolation instead of HAR for dimensionality reduction (ablation) | Hyper-HRNet · HSOD-BIT-V2 | AUC: 0.988 -> 0.728 | Hurt |
| Use PCA instead of HAR for dimensionality reduction (ablation) | Hyper-HRNet · HSOD-BIT-V2 | AUC: 0.988 -> 0.754 | Hurt |
| Use convolution instead of HAR for dimensionality reduction (ablation) | Hyper-HRNet · HSOD-BIT-V2 | AUC: 0.988 -> 0.833 | Hurt |
| Remove dimensionality reduction (ablation) | Hyper-HRNet · HSOD-BIT-V2 | AUC: 0.664 -> 0.988 | Hurt |
| Add MSSA (ablation) | Hyper-HRNet · HSOD-BIT-V2 | AUC: 0.832 -> 0.988 | Helped |
| Add ASAM (ablation) | Hyper-HRNet · HSOD-BIT-V2 | AUC: 0.822 -> 0.988 | Helped |
Surprising findings
- RGB-based U2Net nearly matches Hyper-HRNet on the material similarity attribute, where HSI methods are expected to have an advantage.
- Performance on HSOD-BIT-V2 is substantially lower than on HSOD-BIT, reflecting its greater difficulty.
Limitations
- Hyper-HRNet has a larger parameter count than the RGB baselines due to HRNet.
- Hyper-HRNet-Lite underperforms SMN on some metrics such as MAE on HSOD-BIT-V2.
- Material similarity attribute remains difficult, with Hyper-HRNet only nearly matching U2Net.
Experimental setup
Datasets: HSOD-BIT-V2 HSOD-BIT HS-SOD
Benchmarks: HSOD-BIT-V2 HSOD-BIT HS-SOD
Metrics: MAE precision recall avgF1 AUC CC parameters FLOPs
Baselines: Itti BASNet U2Net SelfReformer SAD SED SG SED-SAD SED-SG SUDF SMN DMSSN
Builds on: HRNet Lite-HRNet ViT SMN DMSSN
Closest papers by meaning
- HyperCOD: The First Challenging Benchmark and Baseline for Hyperspectral Camouflaged Object Detection (AAAI 2026 · similarity 0.89)
- DVSOD: RGB-D Video Salient Object Detection (NeurIPS 2023 · similarity 0.89)
- Alignment-Free RGB-T Salient Object Detection: A Large-Scale Dataset and Progressive Correlation Network (AAAI 2025 · similarity 0.87)
- DiMSOD: A Diffusion-Based Framework for Multi-Modal Salient Object Detection (AAAI 2025 · similarity 0.86)
- M$^{2}$SODAI: Multi-Modal Maritime Object Detection Dataset With RGB and Hyperspectral Image Sensors (NeurIPS 2023 · similarity 0.86)
- ESNet: Evolution and Succession Network for High-Resolution Salient Object Detection (ICML 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.