Scalpel-SAM: Semi-Supervised Infrared Object Detection
Research Paper#Computer Vision, Object Detection, Semi-Supervised Learning, Infrared Imaging🔬 Research|Analyzed: Jan 3, 2026 16:27•
Published: Dec 27, 2025 05:59
•1 min read
•ArXivAnalysis
This paper addresses the critical problem of data scarcity in infrared small object detection (IR-SOT) by proposing a semi-supervised approach leveraging SAM (Segment Anything Model). The core contribution lies in a novel two-stage paradigm using a Hierarchical MoE Adapter to distill knowledge from SAM and transfer it to lightweight downstream models. This is significant because it tackles the high annotation cost in IR-SOT and demonstrates performance comparable to or exceeding fully supervised methods with minimal annotations.
Key Takeaways
- •Addresses data scarcity in IR-SOT using a semi-supervised approach.
- •Leverages SAM as a teacher model.
- •Proposes a two-stage paradigm: Prior-Guided Knowledge Distillation and Deployment-Oriented Knowledge Transfer.
- •Employs a Hierarchical MoE Adapter.
- •Achieves performance comparable to or surpassing fully supervised methods with minimal annotations.
Reference / Citation
View Original"Experiments demonstrate that with minimal annotations, our paradigm enables downstream models to achieve performance comparable to, or even surpassing, their fully supervised counterparts."