Ferkans — Interactive Telecom Tutor

Reproducible RF Imaging Research

Reproducibility is the cornerstone of scientific progress. In RF imaging, reproducibility requires sharing: (1) the data or simulation code; (2) the algorithm implementation; (3) the evaluation pipeline; and (4) the exact parameters used.

The community has made significant progress in open-source tools and standardised benchmarks, but challenges remain --- especially for hardware-in-the-loop experiments. This section surveys the key software libraries and provides a recommended pipeline.

Definition:
DeepInverse Library

DeepInverse (Tachella et al., 2023) is a PyTorch library for solving imaging inverse problems with deep learning:

Forward operators: blur, inpainting, compressed sensing, MRI, tomography. Custom operators easily added.
Reconstruction methods: PnP-ADMM, RED, unrolled networks, diffusion posterior sampling, equivariant imaging.
Training: supervised, self-supervised (Noise2Noise, SURE, equivariant), and unsupervised (DIP).
Evaluation: PSNR, SSIM, LPIPS built-in.
Pretrained denoisers: DnCNN, DRUNet, SwinIR available.

For RF imaging: define the sensing matrix $\mathbf{A}$ as a custom LinearPhysics operator and use any built-in reconstruction algorithm.

Definition:
ODL (Operator Discretization Library)

ODL provides an abstract framework for inverse problems based on operator theory:

Operators: define forward operators as Operator objects with domain, range, and adjoint.
Spaces: discretised function spaces with inner products.
Solvers: Landweber, conjugate gradient, ADMM, Douglas-Rachford.
Integration: interfaces with ASTRA (tomography), PyFFTW, and scikit-image.

ODL's operator abstraction aligns well with the sensing matrix $\mathbf{A}$ formalism used throughout this book: define $\mathbf{A}$ as an Operator and its adjoint $\mathbf{A}^{H}$ follows automatically.

Definition:
Pyxu and PyVista

Pyxu (formerly Pycsou) is a Python library for proximal algorithms:

Proximal operators for $\ell_1$ , TV, nuclear norm, etc.
Splitting algorithms: ADMM, primal-dual, forward-backward.
GPU support via CuPy.
Composable operator framework.

PyVista is a 3D visualisation library built on VTK:

Renders voxel grids, point clouds, and meshes.
Useful for visualising 3D RF imaging reconstructions.
Supports isosurface extraction, volume rendering, and comparison of ground truth vs. reconstruction.

Definition:
Reproducibility Checklist for RF Imaging

A reproducible RF imaging experiment should provide:

Data: raw measurements or simulation code with fixed seeds.
Code: full implementation of the reconstruction algorithm, including pre-processing and post-processing.
Environment: software versions (Python, PyTorch, CUDA) and hardware specifications (GPU model, RAM).
Parameters: all hyperparameters (regularisation $\lambda$ , step sizes, network architecture, training epochs).
Metrics: evaluation code that computes all reported metrics.
Baselines: implementation of comparison methods under identical conditions.

Publishing code on GitHub and data on Zenodo or IEEE DataPort enables independent verification. Many top venues now require code availability for publication.

Recommended End-to-End Evaluation Pipeline

Complexity:

O(N_{\mathrm{MC}} \times K \times T_{\mathrm{recon}})

where

T_{\mathrm{recon}}

is the slowest algorithm's time.

INPUT: Scene distribution P, algorithms A_1..A_K, N_MC trials

OUTPUT: Metric table with confidence intervals

1. FIX random master seed S_0

2. FOR n = 1, ..., N_MC:

a. SAMPLE scene c ~ P using seed S_0 + n

b. GENERATE measurements y = A_fwd * c + noise

(A_fwd differs from reconstruction model)

c. FOR each algorithm A_k:

i. RECONSTRUCT c_hat_k = A_k(y)

ii. COMPUTE metrics: PSNR_k, SSIM_k, P_D_k, time_k

d. STORE results

3. AGGREGATE statistics:

a. Mean +/- 95% CI for each metric

b. Paired t-test for each pair (A_i, A_j)

c. Cohen's d effect sizes

4. GENERATE figures:

a. Metric vs SNR curves with error bars

b. Example reconstructions (best, median, worst)

c. ROC curves

5. PUBLISH: code, data, seeds, and the pipeline script

The pipeline should be a single executable script that produces all results from scratch. No manual intervention or cherry-picking.

,

Software Libraries for RF Imaging

Library	Focus	Forward Model	Reconstruction	GPU	RF Imaging Ready
DeepInverse	Deep inverse problems	Custom operators	PnP, unrolling, diffusion	Yes	Needs custom A
ODL	Operator framework	Custom operators	Landweber, CG, ADMM	Limited	Needs custom A
Pyxu	Proximal algorithms	Custom operators	ADMM, PD, FB	Yes (CuPy)	Needs custom A
PyVista	3D visualisation	N/A	N/A	N/A	Visualisation only
SciPy/NumPy	General scientific	Manual	lstsq, fft, optim	No	Full manual setup
PyTorch/TF	Deep learning	Custom	Any learned method	Yes	Full manual setup

Example: Designing a Reproducible CS Imaging Experiment

Design a fully reproducible experiment comparing matched filter, LASSO, and a U-Net for RF image reconstruction. Specify the data, code, and evaluation pipeline.

Solution

Data

Use the CommIT simulator to generate 1000 2D scenes ( $64 \times 64$ grid) from ShapeNet silhouettes. Forward model: Born approximation on a $256 \times 256$ grid (avoids inverse crime). Add AWGN at SNR $\in \{5, 10, 15, 20, 25, 30\}$ dB. Fix master seed to 42.

Algorithms

MF: $\hat{\mathbf{c}}^{\text{BP}} = \mathbf{A}^{H} \mathbf{y}$ (NumPy). LASSO: CVXPY with $\lambda$ chosen by 5-fold cross-validation. U-Net: PyTorch, trained on 800 samples, validated on 100, tested on 100. Architecture: 4 encoder/decoder levels, 64 base channels.

Evaluation

Compute PSNR, SSIM, NMSE, $P_D$ at $P_{\mathrm{FA}} = 10^{-4}$ on the 100-sample test set. Report mean $\pm$ 95% CI. Paired $t$ -tests for all method pairs. Share all code on GitHub with installation instructions. $\square$

,

🎓CommIT Contribution(2026)

CommIT RF Imaging Simulator

A. Rezaei, G. Caire — TU Berlin CommIT Group, Technical Report

The CommIT group has developed a modular Python simulator for RF imaging that implements the Kronecker-structured sensing model $\mathbf{A} = \mathbf{A}_{\mathrm{freq}} \otimes \mathbf{A}_{\mathrm{space}}$ with GPU acceleration via CuPy and PyTorch. The simulator supports the full 13-phase pipeline from scene generation through Monte Carlo evaluation, and has been used for all experimental results in this book. The emphasis on geometric consistency in the forward model distinguishes it from standard channel simulators.

simulatorKroneckerGPUMonte Carlo

🔧Engineering Note

Publishing Measurement Data

Sharing raw measurement data enables reproducibility and benchmarking. Best practices:

Format: HDF5 or MATLAB .mat with structured metadata (antenna positions, frequencies, timestamps, calibration).
Documentation: include a README with measurement setup, procedure, and known issues.
Licence: use a permissive licence (CC-BY or MIT) to encourage reuse.
Repository: archive on Zenodo, IEEE DataPort, or a university data repository for persistent DOI.
Baseline code: include code that loads the data and produces a basic image (matched filter or backpropagation).

Common Mistake: Unfair Algorithm Comparisons

Mistake:

Comparing a well-tuned deep learning method against an untuned classical baseline (e.g., LASSO with a suboptimal $\lambda$ ).

Correction:

Ensure all methods are given their best chance: (1) Tune hyperparameters for each method using the validation set. (2) Use the same training/validation/test split for all methods. (3) Report the tuning procedure and final hyperparameter values. (4) Use established implementations when available (not custom reimplementations that may contain bugs). (5) Report computational cost (time, memory) alongside accuracy.

Historical Note: The Reproducibility Crisis in Science

2010s

A 2016 Nature survey found that 70% of researchers had failed to reproduce another scientist's experiments, and over 50% had failed to reproduce their own. Signal processing and machine learning are not immune: a 2019 study found that fewer than 30% of NeurIPS papers provided sufficient code to reproduce results. The RF imaging community is still young enough to establish strong reproducibility norms from the start.

DeepInverse

A PyTorch library for solving imaging inverse problems with deep learning, supporting PnP, unrolling, diffusion, and self-supervised training methods.

ODL (Operator Discretization Library)

A Python framework for inverse problems based on abstract operator theory, providing operator composition, adjoint computation, and iterative solvers.

Quick Check

You want to implement PnP-ADMM for RF imaging with a pretrained DRUNet denoiser. Which library provides this out of the box?

ODL

DeepInverse

Pyxu

RadarSimPy

Correction:

DeepInverse

Correct. DeepInverse provides PnP-ADMM with pretrained DRUNet, DnCNN, and SwinIR.

Key Takeaway

DeepInverse provides ready-made PnP, unrolling, and diffusion reconstruction with pretrained denoisers. ODL and Pyxu offer abstract operator frameworks for classical algorithms. PyVista enables 3D visualisation. Reproducibility requires sharing code, data, seeds, and the complete evaluation pipeline. Fair comparisons demand equal tuning effort for all methods and identical evaluation conditions.

Software Libraries and Reproducibility

Reproducible RF Imaging Research

Definition: DeepInverse Library

Definition: ODL (Operator Discretization Library)

Definition: Pyxu and PyVista

Definition: Reproducibility Checklist for RF Imaging

Recommended End-to-End Evaluation Pipeline

Software Libraries for RF Imaging

Example: Designing a Reproducible CS Imaging Experiment

Data

Algorithms

Evaluation

CommIT RF Imaging Simulator

Publishing Measurement Data

Common Mistake: Unfair Algorithm Comparisons

Historical Note: The Reproducibility Crisis in Science

DeepInverse

ODL (Operator Discretization Library)

Quick Check

Key Takeaway

Definition:
DeepInverse Library

Definition:
ODL (Operator Discretization Library)

Definition:
Pyxu and PyVista

Definition:
Reproducibility Checklist for RF Imaging