Chapter Summary

Key Points

1.
The sim-to-real gap (10--15 dB PSNR degradation) is the central challenge for deploying learned RF imaging. Domain adaptation strategies (randomisation, fine-tuning, adversarial, self-supervised) partially address it, but quantifying the gap a priori and guaranteeing worst-case robustness remain open problems. The most effective current approach is tiered simulation combined with self-supervised fine-tuning on a small real dataset.
2.
Dynamic scene imaging requires temporal priors (smoothness, sparse innovation, optical flow, Kalman filtering, learned models) to compensate for per-frame underdetermination. 4D neural fields extending NeRF/3DGS to time are promising but data-intensive; ISAC resource allocation for dynamic imaging is largely unexplored.
3.
Primitive-based scene representations exploit the geometric structure of indoor environments, achieving compression ratios exceeding $10{,}000\times$ over voxel grids. By decomposing scenes into boxes, cylinders, and planes with complex reflectivities, the approach enables physically interpretable reconstructions and natural BIM/CAD integration. This is a largely open research direction with significant potential.
4.
Foundation models for RF face data scarcity and configuration diversity challenges. Cross-modal pre-training (optical to RF) is the most promising near-term approach. RF foundation models will likely be smaller and more physics-aware than their optical counterparts.
5.
Scalability to campus- and city-scale scenes requires hierarchical representations (octrees, block-NeRFs, tiled 3DGS) that reduce memory from $\mathcal{O}(N^3)$ to $\mathcal{O}(N^2)$ . Block boundary consistency, distributed training, and multi-frequency support are unsolved challenges.
6.
Theoretical frontiers include sample complexity of learned imaging, resolution limits with learned priors, the imaging capacity framework connecting resolution to channel capacity, and generalisation bounds for physics-informed networks. The imaging capacity $C = \log\det(\mathbf{I} + \text{SNR}\cdot\mathbf{A}^H\mathbf{A}/N)$ reveals a deep connection between resolvable scene degrees of freedom and MIMO channel capacity.
7.
Reading and writing RF imaging papers requires critical evaluation of assumptions (the assumption audit), verification of claims against evidence, and awareness of the inverse crime. Good papers include fair tuned baselines, statistical rigour, ablation studies, and code/data availability.

Looking Ahead

This chapter -- and this book -- conclude here. The field of RF imaging is at an inflection point: the convergence of 6G standardisation, neural scene representations, and integrated sensing-communication creates unprecedented opportunities for learned imaging methods. The open problems surveyed in this chapter will shape the research agenda for the coming decade. We hope that the foundations laid in Parts I through IX equip the reader to contribute to this agenda. The tools are in your hands; the problems are waiting.

Reading and Writing RF Imaging Papers Exercises