Rf Localization Eval

rf-localization-eval

Radiance-Field Reinforced Pretraining: Scaling Localization Models with Unlabeled Wireless Signals — Wang et al. (2025) (arXiv:2512.07309, 2025)

What this evaluates

Evaluates indoor wireless transmitter localization accuracy using spatial spectrum inputs. It probes the model's ability to learn scene-agnostic spatial-spectral representations from unlabeled RF data and generalize across diverse indoor environments.

Datasets

• Indoor RF Localization Scenes — total ?; splits: test (25)

Metrics

• Euclidean distance (cm) (primary) — range: cm
  • Euclidean distance between the predicted and ground-truth transmitter positions, reported in centimeters. Lower values indicate better localization accuracy.
• SSIM — range: [0, 1]
  • Structural Similarity Index measuring reconstruction quality between original and predicted spatial spectra. Ranges from 0 to 1, where higher values indicate better preservation of spectral structure.

Input / output format

Input: Standardized 36×9 spatial spectrum matrix representing RF signal data.

Output: Predicted transmitter position coordinates.

Scoring recipe

def compute_metrics(predictions, ground_truth):
    # predictions and ground_truth are arrays of shape (N, spatial_dims)
    errors = np.linalg.norm(predictions - ground_truth, axis=1)
    mean_error_cm = np.mean(errors) * 100  # Convert to cm if inputs are in meters
    return mean_error_cm

Common pitfalls

• Results vary significantly based on the proportion of labeled fine-tuning data (20%–80%).
• Pretraining masking ratio must be tuned (optimal ~75%) to avoid information loss or insufficient regularization.
• Models must be adapted to a standardized 36×9 spatial spectrum input for fair comparison.

Evidence (verbatim from paper)

Localization accuracy is measured by the Euclidean distance between the predicted and ground-truth transmitter positions.

Citation

@misc{wang2025radiance,
  title={Radiance-Field Reinforced Pretraining: Scaling Localization Models with Unlabeled Wireless Signals},
  author={Wang et al. (2025)},
  year={2025},
  note={arXiv:2512.07309}
}

• arXiv: 2512.07309