Reproduce “Temporal-Emerged Prompting for Segment Anything in Multiframe Infrared Small Target Detection” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yinghui Xing, Donghao Chu, Shizhou Zhang, Di Xu
TEP-SAM, a framework developed by Northwestern Polytechnical University and Huawei Technologies Ltd., integrates temporal-emerged cues into the Segment Anything Model (SAM) for multiframe infrared small target detection. The method enables superior localization and segmentation of dim targets in low signal-to-noise ratio sequences, consistently outperforming existing approaches with IoU improvements of up to 10.07 points on challenging datasets.
Key insights
- Despite being indistinct in single frames, small targets in infrared sequences gradually emerge and become distinguishable from background dynamics over time, providing a crucial discriminative cue.
- Explicitly leveraging these temporal-emerged cues through both feature injection into the encoder and automatic prompt generation for the decoder effectively adapts large-scale foundation models like SAM to specialized, low-contrast M-IRSTD tasks.
- The combination of global temporal modeling, local motion analysis, and temporal discrepancy modulation within DETE provides complementary information, enhancing target features and suppressing background clutter for robust segmentation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.