Datasets for RF Imaging Research

ShapeNet: a large repository of 3D CAD models ($> 51{,}000$ models in 55 categories). Used to generate diverse target geometries for training learned reconstruction methods. Each model can be voxelised onto the imaging grid to produce a ground-truth reflectivity map.

THuman: a collection of 3D human body meshes from body scanning. Used for human-body RF imaging (e.g., through-wall sensing, vital sign monitoring) where the target has realistic human anatomy and posture.

MPEG-7 Shape Dataset: 2D shape silhouettes used as targets for 2D RF imaging experiments. Simpler than 3D models but useful for rapid prototyping and algorithm comparison.

Automotive radar datasets:

• nuScenes: 1000 driving scenes with 5 radar sensors, camera, and LiDAR. 3D bounding boxes for 23 object classes.
• RadarScenes: 158 sequences, 4 hours of driving data with radar point clouds and annotations.

Indoor imaging datasets:

• DeepMIMO: simulated massive MIMO channel dataset from ray tracing, covering multiple indoor and outdoor scenarios at 2.4, 28, and 60 GHz.
• DeepSense 6G: multi-modal sensing dataset for 6G research, including radar, camera, LiDAR, and GPS data.

SAR datasets:

• MSTAR: Moving and Stationary Target Recognition dataset. SAR images of military vehicles at various aspect angles.
• SEN12MS: multi-modal satellite dataset (SAR + optical) for land use classification.

When real data is unavailable or insufficient, simulation provides controlled, labelled data. A synthetic data pipeline has four stages:

1. Scene sampling: draw a random scene from a distribution (e.g., sample a ShapeNet model, place it at random position/orientation, set reflectivity parameters).

2. Forward model: compute synthetic measurements $\mathbf{y} = \mathbf{A}\mathbf{c} + \mathbf{w}$ using the chosen physics (Born model, ray tracing, or full-wave).

3. Noise and impairments: add receiver noise ($\mathbf{w} \sim \mathcal{CN}(0, \sigma^2\mathbf{I})$), phase noise, clutter, and hardware impairments.

4. Ground truth pairing: store both the measurement $\mathbf{y}$ and the scene $\mathbf{c}$ as a (input, target) pair for supervised training.

The critical question: does the forward model used for data generation match the one used for reconstruction? If yes, you have committed the inverse crime.

The inverse crime is committed when the same forward model generates the synthetic data and reconstructs from it:

$$\hat{\mathbf{c}} = \arg\min_{\mathbf{c}} \|\mathbf{A}_{\mathrm{inv}} \mathbf{c} - \underbrace{\mathbf{A}_{\mathrm{fwd}} \mathbf{c}_{\mathrm{true}} + \mathbf{w}}_{\text{synthetic data}}\|^2 \qquad \text{with } \mathbf{A}_{\mathrm{inv}} = \mathbf{A}_{\mathrm{fwd}}.$$

When $\mathbf{A}_{\mathrm{fwd}} = \mathbf{A}_{\mathrm{inv}}$, the reconstruction problem reduces to denoising: the model mismatch error is zero, and the only error source is noise. This makes any reconstruction method look artificially good.

The crime is aggravated when:

• The same discretisation grid is used for generation and reconstruction;
• Point scatterers are placed exactly on grid points;
• The same random seed generates training and test data.

To avoid the inverse crime, ensure the forward model used for data generation differs meaningfully from the reconstruction model:

1. Different discretisations: generate data on a fine grid ($N_{\mathrm{fwd}} \gg N_{\mathrm{inv}}$) and reconstruct on a coarser grid. Rule of thumb: $N_{\mathrm{fwd}} \geq 4 N_{\mathrm{inv}}$.

2. Different physics: generate with ray tracing or FDTD; reconstruct with the Born approximation.

3. Off-grid scatterers: place point targets at positions that do not coincide with the reconstruction grid.

4. Model mismatch injection: add calibration errors, timing offsets, and mutual coupling to the simulated data but not to the reconstruction model.

5. Real data validation: always validate on measured data when available.