AIBRIEF
Posted as arXiv preprint 2607.24184 and released on GitHub, LCMamNet, a model built to spot faint heat signatures in infrared images, reports 71.25–95.58% mIoU (mean intersection over union) on three public infrared benchmarks.
A 1.18-million-parameter AI model spots faint infrared targets in 6.62 milliseconds on a Jetson edge device, fast enough to track heat signatures from a single onboard computer without server-side help.
The model, called LCMamNet, runs on an NVIDIA Jetson Orin NX 16G SUPER and was posted this week as arXiv preprint 2607.24184, with code and checkpoints on GitHub. At a 256×256 input size, the network uses 6.91 billion floating-point operations per frame, leaving headroom on a drone or perimeter sensor for other perception tasks.
The target application is infrared small target detection: spotting a drone, a person, or a vehicle whose heat signature covers only a few pixels and is easily lost in cloud edges, ground clutter, or sun glare. The authors report mean intersection-over-union of 71.25% on IRSTD-1k, 79.60% on NUAA-SIRST, and 95.58% on NUDT-SIRST, three public benchmarks for this task.
Every number above is self-reported by the authors on their own preprint. The paper has not been peer-reviewed. There is no independent benchmark yet, no fielded comparison against current production counter-drone or security infrared systems, and no data on how the model behaves against birds, weather, or sun glare beyond the single Jetson experiment.