Skip to content

Dataset

This dataset provides height maps, configuration-only beam maps, and ray-tracing radiomaps for studying multi-configuration radiomap prediction in U6G / XL-MIMO systems.

It is designed to support research on:

  • cross-configuration generalization
  • cross-environment generalization
  • beam-aware radiomap prediction
  • sparse radiomap reconstruction

The released benchmark separates three complementary data components:

  • height maps for scene geometry representation
  • beam maps for configuration-aware and environment-independent side information
  • radiomaps for environment-dependent ray-tracing labels

Visual Examples

Representative examples from the released dataset are shown below.

Overall Preview

An overall preview of the released data components, including the height map, ray-tracing radiomap, and configuration-only beam map.

Radiomap Examples

Representative ray-tracing radiomap examples from different scenes and configurations.

Height Map Examples

Representative height map examples from the released urban scenes.

Beam Map Examples

Representative configuration-only beam map examples.

Paired Examples

The following examples illustrate the correspondence among the released data components for selected scenes and transmitter configurations.

Dataset Statistics

Item Value
Scenes 800
Configurations 98
Frequencies 1.8 / 2.6 / 3.5 / 4.9 / 6.7 GHz
Transmit antennas 4 → 1024 TR
Beam counts 1 / 8 / 16 / 64
Beam pattern 3GPP TR 38.901

---

Folder Structure

Dataset/
  beam_maps/
    <config_id>/
      u0/
        beam_XX_angle_*.npy
        u0_all_beams.npz
        beam_settings.txt
        *_plot.png (optional)
  configs/
    *.txt
  height_maps/
    u1..u800/
      u*_height_matrix.npy
      u*_height_matrix_coords.npz
  radiomaps/
    <config_id>_beamXX/
      u1_labeled_radiomap.npy
      ...
      u800_labeled_radiomap.npy
      beam_settings.txt
  sionna_maps/ (optional)
    u1..u800/
      meshes/*.ply

Dataset Components

1. Height Maps

Height maps provide geometry-aware scene information for each environment.

Stored under:

height_maps/u*/

Each scene contains:

  • u*_height_matrix.npy
  • u*_height_matrix_coords.npz

Height matrix

u*_height_matrix.npy

A 2.5D height map describing building and terrain height on a regular grid.

Coordinate metadata

u*_height_matrix_coords.npz

Stores the grid-coordinate metadata associated with the height matrix.

Default generation settings

  • target area: 1280 m × 1280 m
  • grid resolution: 5 m
  • default height per floor: 5 m
  • single-floor buildings are assigned a height of 20 m

Notes

  • height maps are scene-dependent
  • they are used as geometry-related inputs for radiomap learning
  • if multiple buildings overlap on the same grid cell, the maximum height is used

2. Beam Maps (Configuration-Only)

Beam maps provide configuration-aware but environment-independent side information.

Stored under:

beam_maps/<config_id>/u0/

Important

u0 does not correspond to a real scene.

It is a placeholder used to store configuration-only beam map features that depend on transmitter configuration and beam index, but not on a specific environment. The same beam maps can therefore be reused across all scenes.

Typical files

Per-beam matrices

beam_XX_angle_*.npy

Per-beam matrix representations of the beam pattern.

All-beam package

u0_all_beams.npz

Compressed package containing all beam maps under one configuration.

Beam metadata

beam_settings.txt

Beam and configuration metadata for the generated beam maps.

Visualization (optional)

*_plot.png

Visualization of beam patterns.

Modeling assumptions

The released beam maps are generated from a simplified geometric model:

  • beam maps are not ray-tracing radiomaps
  • only LOS geometry is considered
  • no reflection, diffraction, blockage, or occlusion is modeled

The released generator supports:

  • plane-wave approximation
  • spherical-wave model

Practical usage rule

When predicting radiomaps for configuration <config_id>, always load beam maps from:

beam_maps/<config_id>/u0/

Beam maps should be paired with radiomaps using the same configuration and same beam index.

3. Radiomaps (Labels)

Radiomaps provide the environment-dependent labels generated by the ray-tracing pipeline.

Stored under:

radiomaps/<config_id>_beamXX/

Each folder corresponds to one configuration + one beam.

Files

Radiomap label

u*_labeled_radiomap.npy

Radiomap corresponding to:

  • scene u*
  • configuration <config_id>
  • beam beamXX

Beam metadata

beam_settings.txt

Configuration metadata duplicated for convenience.

Example access

Scene:

u123

Configuration:

freq_6.7GHz_256TR_16beams_pattern_tr38901

Beam:

beam03

Radiomap file:

Dataset/radiomaps/freq_6.7GHz_256TR_16beams_pattern_tr38901_beam03/u123_labeled_radiomap.npy

Label semantics

The released labeled radiomap uses the following conventions:

  • building region: 1000
  • invalid / no-label region: -300
  • valid propagation region: values strictly between -300 and 0 dB

This means:

  • building cells are explicitly marked in the labeled radiomap
  • invalid cells correspond to the ray-tracing floor value
  • only valid propagation cells represent effective supervision targets in downstream preprocessing

Notes

  • radiomaps are environment-dependent ray-tracing labels
  • under the default released generation settings, invalid cells correspond to -300 dB
  • building regions are explicitly marked by 1000 in the labeled radiomap

Indexing and Naming Rules

Scene ID (u1..u800)

u1..u800 denote different scenes / environments.

Each scene corresponds to a distinct geographic region with its own height map and ray-tracing environment.

Configuration ID (<config_id>)

Configuration folders follow the naming format:

freq_{f}GHz_{NTR}TR_{B}beams_pattern_tr38901

Examples:

freq_1.8GHz_4TR_1beams_pattern_tr38901
freq_6.7GHz_1024TR_64beams_pattern_tr38901

Interpretation

Field Meaning
f carrier frequency (GHz)
NTR number of transmit antennas
B number of beams in the codebook
pattern_tr38901 beam pattern follows 3GPP TR 38.901

Beam ID (beamXX)

Beam indices follow:

beam00 .. beam{B-1}

where B is the beam count of the configuration.

Examples

Configuration Beam IDs
1 beam beam00
8 beams beam00 .. beam07
16 beams beam00 .. beam15
64 beams beam00 .. beam63

Mesh Assets (Optional)

The directory

sionna_maps/

contains .ply meshes used to reproduce the ray-tracing scenes.

Example structure:

sionna_maps/
  u1/
    meshes/*.ply

Usage

Mesh assets are required only if reproducing the ray-tracing pipeline.

They are not required for training or evaluating machine learning models using the released height maps, beam maps, and radiomaps.

Summary of Roles

The three released data components play different roles:

Component Depends on scene? Depends on configuration? Main role
Height map Yes No geometry-aware scene input
Beam map (u0) No Yes configuration-aware side information
Radiomap Yes Yes environment-dependent learning target

This separation is one of the key design choices of the released benchmark.

Used by the Baselines

The released baseline pipelines use these dataset components in different but coordinated ways:

  • GAN and UNet preprocessing both use the height map as geometry-related input
  • feature-map mode loads beam maps from the configuration-only reference folder u0
  • radiomap preprocessing extracts the building mask and valid mask from the labeled radiomaps

In particular:

  • the building region is identified from the labeled radiomap value 1000
  • the invalid / no-label region is identified from the default floor value -300
  • only valid propagation regions are used as effective supervision targets in downstream preprocessing.