Chapter Summary

Chapter Summary

Key Points

• 1.

  Hardware platforms for RF imaging span SDR boards (flexible, sub-6 GHz), commercial mmWave modules (low-cost, 60--79 GHz, MIMO), and channel sounders (highest precision). The choice depends on the research goal: prototyping, demonstration, or ground-truth measurement. Calibration (phase, timing, amplitude, mutual coupling) is essential regardless of platform.

• 2.

  Standard datasets (nuScenes, MSTAR, DeepMIMO, ShapeNet, THuman) provide benchmarking data, but raw signal-level data remains scarce. Synthetic generation via ray tracing or Born-model simulation fills the gap, but demands vigilance against the inverse crime: use different grids ($4\times$ oversampling), different physics, off-grid targets, and real-data validation.

• 3.

  Simulation frameworks (Sionna RT, FEKO, RadarSimPy) cover different fidelity levels. The CommIT simulator uses Kronecker-structured sensing for fast GPU-accelerated 2D/3D imaging with a 13-phase pipeline. Monte Carlo evaluation with $N_{\mathrm{MC}} \geq 100$ trials, confidence intervals, and paired statistical tests is essential for credible claims.

• 4.

  Evaluation metrics include image quality (PSNR, SSIM, LPIPS, NMSE), shape quality (Chamfer, Hausdorff, IoU, F-score), detection ($P_D$, $P_{\mathrm{FA}}$, ROC, AUC), and computational cost (FLOPs, time, memory). No single metric tells the full story --- report multiple metrics. PSNR can disagree with SSIM and detection metrics; task-specific metrics are most meaningful for end applications.

• 5.

  Software libraries (DeepInverse, ODL, Pyxu, PyVista) provide modular tools for reproducible research. A complete pipeline from data generation through reconstruction and evaluation, shared as executable code with fixed seeds and documented parameters, is the standard for credible publications.