Ximeng Zhai, Zheng Wang, Yaohong Chen +3 more
Optics has a hard limit on how finely it can resolve detail, and over long distances two nearby infrared sources stop being two things in the image and become one indistinct blob. Their energy overlaps. Ordinary detection assumes each thing in the world maps to one thing in the picture, and that assumption has quietly failed.
The response is to stop detecting and start unmixing: take the blob and decompose it back into the separate sub-targets that produced it. Think of two distant headlights merging into a single smear, where the task is no longer to spot the smear but to work out that it is two lights.
The dominant way of doing this, deep unfolding networks, repeats an iterative structure many times over. That inheritance brings high latency and a rigid architecture that is hard to adapt. This paper's contribution is a one-stage alternative built to avoid paying that repetition cost.
Due to the optical diffraction limit and long imaging distances, Closely-Spaced Infrared Small Targets (CSIST) typically exhibit energy overlap, manifesting as indistinguishable blobs in infrared images. This ambiguity invalidates the one-to-one mapping assumption of traditional detection, thereby necessitating a paradigm shift towards CSIST Unmixing, which decomposes these blobs into discrete sub-targets. However, the dominant paradigm deep unfolding networks are shackled…
Partition, Prompt, Aggregate: Statistical Self-Consistency in Language Models
arXiv (cs.AI) · July 16, 2026RoboTTT: Context Scaling for Robot Policies
arXiv (cs.LG) · July 16, 2026MeanFlowNFT: Bringing Forward-Process RL to Average-Velocity Generators
arXiv (cs.AI) · July 16, 2026SciDiagramEdit: Learning to Edit Scientific Diagrams from Paper Revisions