Lighted-SAM: Lightening Open-World SAM for Low-Light Segmentation.

Area of Science:

• Computer Vision
• Artificial Intelligence

Background:

• The Segment Anything Model (SAM) demonstrates strong open-world segmentation capabilities.
• SAM's performance degrades significantly in low-light conditions due to underrepresentation in its pre-training dataset (SA-1B).
• Existing methods often focus on specific tasks or closed-set scenarios, limiting applicability to real-world dark environments.

Purpose of the Study:

• To improve SAM's segmentation performance in low-light conditions while maintaining its zero-shot generalization.
• To develop a method applicable to open-world models operating in challenging lighting environments.
• To introduce a novel dataset and a lightweight repair strategy for low-light image segmentation.

Main Methods:

• Construction of DarkSeg58K, a diverse and realistic dataset for low-light segmentation.
• Introduction of Lighted-SAM, a lightweight repair strategy utilizing the Spectral Information Resonance (SIR) mechanism.
• Integration of a spectral enhancement module into a frozen SAM without end-to-end retraining.

Main Results:

• Lighted-SAM effectively enhances SAM's segmentation accuracy in low-light conditions.
• The proposed Spectral Information Resonance (SIR) mechanism facilitates seamless integration of spectral enhancement.
• The approach preserves SAM's robust zero-shot segmentation capabilities.
• Experimental validation on multiple benchmarks confirms the method's superiority.

Conclusions:

• Lighted-SAM offers an effective and efficient solution for improving SAM's low-light performance.
• The DarkSeg58K dataset provides valuable resources for research in low-light computer vision.
• This work advances the practical application of powerful segmentation models in challenging real-world scenarios.