MilliSARImageNet: A 2D High-Resolution Millimeter-Wave SAR Image Dataset

Millimeter-wave (mmWave) imaging is a key sensing modality for cyber-physical systems (CPS), enabling applications in autonomous robotics, industrial inspection, and security screening. Yet collecting large, labeled mmWave imaging datasets is costly, limiting progress in robust perception. We present MilliSARImageNet, a high-resolution mmWave imaging dataset that couples a physics-based digital-twin simulation with real measurements from a 77–81 GHz imager. The digital twin replicates the real imager’s geometry, waveform, and reconstruction pipeline, enabling scalable, physically consistent data generation, while the real subset provides an evaluation reference for synthetic-to-real transfer. Images are reconstructed using four imaging algorithms, and targeted augmentations are applied to capture realistic variability. Fine-tuning ConvNeXt-B, Swin-B, ViT-B/16, and ResNet-152 solely on synthetic images yields 95%–98% classification accuracy on unseen synthetic images and 81%–88% on unseen real images. These results show that our digital-twin approach is effective while still revealing a remaining synthetic-to-real domain gap. MilliSARImageNet provides a reproducible foundation for studying domain adaptation and robust, trustworthy mmWave perception in CPS.