Multi-Scale Attention Module Enhances Hyperspectral Segmentation for Autonomous Driving
A recent preprint on arXiv (2506.18682) introduces a Multi-Scale Attention Mechanism (MSAM) aimed at improving semantic segmentation for autonomous vehicles through hyperspectral imaging. Although this imaging technique provides intricate spectral information, its high dimensionality poses challenges. MSAM utilizes three parallel 1D convolutional layers with varying kernel sizes between 1 and 11 to effectively capture multi-scale features, which are then adaptively combined for enhanced spectral extraction. Integrated within the UNet framework, MSAM is tested on hyperspectral datasets relevant to urban driving scenarios. Ablation studies indicate that MSAM surpasses the baseline UNet with standard skip connections (UNet-SC), yielding average enhancements of 2.32% in mean Intersection over Union (mIoU) and 2.88% in mean F1-score (mF1), while ensuring competitive GPU efficiency.
Key facts
- Paper arXiv:2506.18682 proposes a Multi-Scale Attention Mechanism (MSAM).
- MSAM uses three parallel 1D convolutions with kernel sizes from 1 to 11.
- The mechanism is integrated into UNet skip connections.
- Evaluation is performed on multiple HSI datasets for urban driving.
- MSAM improves mIoU by an average of 2.32% over baseline UNet-SC.
- MSAM improves mF1 by an average of 2.88% over baseline UNet-SC.
- The method maintains competitive GPU performance.
- The paper is an arXiv preprint announced as type replace-cross.
Entities
Institutions
- arXiv