---
title: '3DiRM3200: 3D Indoor Radio Map Datasets'
url: https://www.emergentmind.com/topics/3d-indoor-radio-map-dataset-3dirm3200
type: topic
---

# 3DiRM3200: 3D Indoor Radio Map Datasets

3D Indoor Radio Map Dataset, abbreviated 3DiRM3200, is a designation used in the arXiv literature for several indoor wireless-propagation datasets rather than for a single unambiguous resource. In one usage, it denotes a multimodal dataset that integrates high-resolution 3D LiDAR scans with Wi-Fi Received Signal Strength Indicator (RSSI) measurements collected under 20 distinct Access Point (AP) configurations in a multi-room indoor environment [2511.00494]. In an earlier usage, it identifies a drone-supported indoor Wi-Fi measurement dataset built from beacon observations at discrete 3D waypoints [2109.06923]. In a later deep-learning benchmark, it refers to a publicly released, fully simulated dataset of indoor radio maps created with WinProp’s Dominant Path Model (DPM) [2605.18127]. The shared name reflects different measurement logics, data modalities, and intended uses, and this naming overlap is central to understanding the term in current technical discourse.

## 1. Nomenclature and scope in the literature

In the cited literature, the designation 3DiRM3200 refers to three distinct resources [2511.00494] [2109.06923] [2605.18127]. A recurrent source of confusion is that the suffix “3200” does not encode the same quantity across these works.

| Source paper | Resource type | Meaning of “3200” |
|---|---|---|
| “A Multimodal Dataset for Indoor Radio Mapping with 3D Point Clouds and RSSI” [2511.00494] | Real multimodal LiDAR + RSSI dataset | Approximately 3.2 million 3D points in the combined office + corridor point cloud |
| “Research Project 2: Drone-supported AI-based Generation of 3D Maps of Indoor Radio Environments” [2109.06923] | Real drone-acquired Wi-Fi waypoint dataset | Nominal, targeted number of measurements: \(72\) scan-locations \(\times \approx 44\) AP receptions per \(3\) s scan \(\simeq 3168\), rounded to \(3200\) |
| “R\(^2\)Net: 2D Deep Residual Learning with Height Embedding for 3D Radio Map Estimation” [2605.18127] | Fully simulated 3D radio-map dataset | \(3\,200\) distinct 3D radio-map samples |

This multiplicity of meanings has methodological consequences. A citation to 3DiRM3200 may refer to raw RSSI observations, LiDAR-aligned multimodal measurements, or dense simulated pathloss tensors. A plausible implication is that any comparative use of the term requires explicit disambiguation by paper title or arXiv identifier.

## 2. Multimodal LiDAR–RSSI dataset in a multi-room indoor environment

The 2025 dataset introduced in “A Multimodal Dataset for Indoor Radio Mapping with 3D Point Clouds and RSSI” combines approximately \(3.2\) million LiDAR points with Wi-Fi RSSI measurements collected over \(20\) AP placements and \(53\) grid locations, yielding \(1\,060\) intended RSSI samples and \(1\,027\) valid RSSI readings after discarding weak or erroneous points [2511.00494]. The physical setting consists of a main office room measuring \(7.58\,\mathrm{m}\) in width and \(11.00\,\mathrm{m}\) in length, subdivided into a wood-surfaced upper zone and tiled lower zone, together with a connecting hallway approximately \(2.5\,\mathrm{m}\) wide and \(32.70\,\mathrm{m}\) long, with two bends.

The dataset uses a unified right-handed Cartesian coordinate system. The origin \((0,0)\) is at the southwest corner of the combined office+corridor footprint, the \(X\) and \(Y\) axes are horizontal and measured in metres, and \(Z\) is measured in metres above the floor. All LiDAR and RSSI locations use this shared frame, which is essential for geometry-aware REM construction.

Its LiDAR modality employs Livox Avia and Velodyne (Ouster) VLP-16 devices. The Avia has a vertical field of view of approximately \(38.4^\circ\), while the VLP-16 has a vertical field of view of \(\pm 15^\circ\) (\(30^\circ\) total); both provide \(360^\circ\) horizontal field of view. Scan frequency is \(10\,\mathrm{Hz}\) for the Avia and up to \(20\,\mathrm{Hz}\) for the VLP-16, with point throughput up to \(240\,000\,\mathrm{pts/s}\) and \(300\,000\,\mathrm{pts/s}\), respectively. Average point density is approximately \(1\,000\)–\(2\,000\,\mathrm{pts/m^2}\) in open areas and higher near scanned surfaces.

The Wi-Fi modality is based on a TP-Link TL-WR841N AP in the \(2.4\,\mathrm{GHz}\) band with \(20\,\mathrm{MHz}\) channel width, two omnidirectional dipole antennas, and a transmit power of \(40\,\mathrm{dBm}\) maximum. The receiver is a Samsung SM-A556B smartphone running Android 14 and using the “Wi-Fi Analyzer” app. RSSI sampling is approximately \(1\,\mathrm{Hz}\) per grid location, with recommended dwell time of at least \(5\,\mathrm{s}\) to average out fast fading. The AP height is \(1.2\,\mathrm{m}\), and the user-equipment measurement height is \(1.5\,\mathrm{m}\).

The \(20\) AP configurations are divided into two scenarios. Scenario 1, corresponding to the empty environment, contains \(12\) placements drawn from the \(53\)-point grid, including corners, near-wall placements, and deep-corridor placements. Scenario 2, corresponding to the occupied environment, contains \(8\) placements, of which \(4\) are reused from Scenario 1 and \(4\) are new points in dense desk clusters or the corridor. All setups use channel \(11\) at \(2\,462\,\mathrm{MHz}\) with \(20\,\mathrm{MHz}\) bandwidth, and each AP location corresponds to one of the \(53\) measurement grid points listed in the RSSI `.csv` and `.h5` metadata. The selected propagation conditions span line-of-sight in the open room, non-line-of-sight through one or more walls, long-range corridor links, and reflection-dominated corridor bends.

## 3. Measurement protocol, spatial registration, and data products

The dataset explicitly separates two measurement scenarios, each with a defined acquisition protocol [2511.00494]. In Scenario A, the empty office condition, the procedure is to place the AP at a designated grid point, hold the smartphone at each of the \(53\) red-dot locations, and record RSSI for at least \(5\,\mathrm{s}\). The grid spacing is \(2\,\mathrm{m}\) in both \(X\) and \(Y\), with one exception at \(1\,\mathrm{m}\) due to furniture. Calibration includes a smartphone factory reset, clearing the scanning app cache, and rebooting the AP between runs. Traversal follows the numeric sequence shown in Figure 6b of the paper in order to minimize walking distance and human error.

Scenario B introduces human presence. It includes \(7\)–\(10\) volunteers performing typical office tasks such as typing, walking, and conversing. Four AP placements are identical to those of Scenario A for direct line-of-sight and non-line-of-sight comparisons, and four additional placements are located in heavily occupied desk clusters or in the corridor mid-section. Movement patterns are described as random pedestrian flow, and volunteers were instructed to avoid standing still at measurement points. This design makes the dataset suitable for studying dynamic environmental effects on wireless signal propagation.

The alignment of geometry and signal measurements is manual rather than fully automated. The office and corridor point clouds were captured separately and then manually registered by selecting at least three correspondence landmarks, such as corners and furniture edges, in Open3D. No fully automated Iterative Closest Point procedure was applied; users supply control points in the interactive toolbox. Manual ICP residuals are typically below \(0.05\,\mathrm{m}\), and the expected localization error for each RSSI sample is at most \(0.10\,\mathrm{m}\), dominated by grid-marker placement and smartphone holding variability.

The point clouds are distributed as `office.ply`, `corridor.ply`, and `combined.ply`, in ASCII and binary PLY, with `.las` availability on request. RSSI measurements are provided both as raw CSV and as HDF5. The CSV columns are `timestamp`, `SSID`, `channel`, `measured frequency (MHz)`, `RSSI (dBm)`, `estimated distance (m)`, `setup ID`, `ap_location_id`, `grid_index`, `X`, `Y`, and `Z`. The HDF5 file contains `/data` as an \(N_{\text{setups}} \times N_{\text{points}}\) RSSI matrix, together with `/setup`, `/ap_locations`, and `/indices`. Distribution is through Zenodo at `https://doi.org/10.5281/zenodo.15791300`, under a CC BY 4.0 license.

## 4. Propagation models, empirical statistics, and REM-oriented use

The multimodal dataset is framed around indoor radio propagation and Radio Environment Map estimation, and the accompanying description provides both standard analytical models and empirical statistics [2511.00494]. The principal path-loss model is the log-distance model

$$
L_{\mathrm{PL}}(d) = L_0 + 10\,n\,\log_{10}(d/d_0) + X_\sigma,
$$

where \(L_0 \approx 40\,\mathrm{dB}\) at reference distance \(d_0 = 1\,\mathrm{m}\), \(n \approx 2.0\) for line-of-sight office conditions, \(n \approx 3.4\) for non-line-of-sight corridor conditions, and \(X_\sigma \sim \mathcal{N}(0,\sigma^2)\) with \(\sigma \approx 4\)–\(6\,\mathrm{dB}\).

A free-space reference is also given through the Friis equation,

$$
P_r(d) = P_t + G_t + G_r - 20\log_{10}(4\pi d/\lambda).
$$

The reported empirical statistics differentiate LOS and NLOS behavior. The LOS mean RSSI at \(5\,\mathrm{m}\) is \(-45\,\mathrm{dBm}\) with \(\sigma \approx 3\,\mathrm{dB}\), while the NLOS mean RSSI at \(10\,\mathrm{m}\) is \(-70\,\mathrm{dBm}\) with \(\sigma \approx 5\,\mathrm{dB}\). Human presence increases path loss by \(2\)–\(6\,\mathrm{dB}\) on average and adds \(\sigma' \approx 2\,\mathrm{dB}\) to fading spread. These measurements directly support analysis of occupancy-induced perturbations in indoor propagation.

The stated use cases are machine-learning-based REM estimation, AP placement optimization, testing of emerging standards such as IEEE 802.11be by applying frequency-dependent scaling to measured \(2.4\,\mathrm{GHz}\) data, and hybrid synthetic/real workflows in which the PLY geometry is embedded in ray tracers such as Mitsuba2 and calibrated against measured RSSI. This suggests a role not only as a static benchmark but also as a bridge between geometry-driven and data-driven radio modeling.

## 5. Drone-supported 3D waypoint dataset using the same designation

An earlier arXiv work, “Research Project 2: Drone-supported AI-based Generation of 3D Maps of Indoor Radio Environments,” uses the same identifier 3DiRM3200 for a different indoor radio dataset [2109.06923]. Here, the raw dataset contains \(2696\) Wi-Fi beacon samples collected at discrete 3D waypoints, reduced to \(2565\) samples after dropping MAC addresses with fewer than \(16\) observations. The “3200” label denotes the nominal target number of measurements, computed as \(72\) scan-locations times approximately \(44\) AP receptions per \(3\)-second scan, giving about \(3168\) and rounded to \(3200\).

The measurement space is a rectangular cuboid of \(3.74\,\mathrm{m} \times 3.20\,\mathrm{m} \times 2.10\,\mathrm{m}\), located in a living room of a multi-storey apartment building. Floor, walls, and ceiling are standard drywall and concrete partitions, and eight UWB LPS anchors at the cuboid corners define the origin at \((0,0,0)\) and provide coordinate accuracy of approximately \(10\,\mathrm{cm}\). Two autonomous Crazyflie 2.1 drones, each equipped with a UWB Loco Positioning Deck and a custom ESP8266 Wi-Fi deck, visit \(36\) predefined 3D waypoints each, for \(72\) unique locations in total. The drones operate sequentially to avoid mutual RF interference.

At each waypoint, the protocol comprises \(4\,\mathrm{s}\) of transit from the previous point and a \(3\,\mathrm{s}\) hovering Wi-Fi scan. During the scan, the \(2.4\,\mathrm{GHz}\) control link is muted and a FreeRTOS `hoverWhileScanning` task maintains position. The radio interface performs IEEE 802.11b/g/n scanning in the \(2.4\,\mathrm{GHz}\) ISM band on channels \(1\), \(6\), and \(11\). The recorded tuple is
\[
\langle \text{timestamp}, x[\mathrm{m}], y[\mathrm{m}], z[\mathrm{m}], \text{ssid}, \text{mac}, \text{channel}, \text{rssi}[\mathrm{dBm}] \rangle.
\]

The dataset is organized as CSV files, one per drone or flown mission, with one row per beacon observation. Preprocessing drops the SSID as redundant, removes MAC addresses with fewer than \(16\) observations, and one-hot encodes MAC and channel categories. The resulting machine-learning-ready dataset contains features \(\{x, y, z, \mathrm{rssi}, \mathrm{mac\_onehot}, \mathrm{channel\_onehot}\}\). The paper also reports gap-filling baselines: a per-AP mean baseline achieves RMSE \(=4.8107\,\mathrm{dBm}\), k-nearest neighbors with \((x,y,z)\) and one-hot MAC achieves RMSE \(=4.5920\,\mathrm{dBm}\), a scaled-MAC variant reaches RMSE \(=4.4186\,\mathrm{dBm}\), kNN per MAC gives RMSE \(=4.4267\,\mathrm{dBm}\), and a Keras neural network with \(16\) hidden nodes gives RMSE \(=4.4870\,\mathrm{dBm}\). In this usage, 3DiRM3200 is a sparse waypoint-level RSSI dataset rather than a dense radio-map tensor.

## 6. Fully simulated 3D radio-map benchmark for R\(^2\)Net

A later and again distinct use of the name appears in “R\(^2\)Net: 2D Deep Residual Learning with Height Embedding for 3D Radio Map Estimation” [2605.18127]. This 3DiRM3200 is a publicly released, fully simulated dataset of indoor radio maps created with WinProp’s Dominant Path Model. It contains \(3\,200\) distinct 3D radio-map samples across \(200\) real-floorplan geometries, each equipped with furniture and up to \(16\) transmitter locations. The paper states that the dataset was created because of the lack of publicly available 3D radio map datasets and that its production took more than \(1{,}000\) labour hours.

Each scene covers a \(20.44\,\mathrm{m} \times 20.44\,\mathrm{m}\) indoor area rasterized to a \(256 \times 256\) grid with \(8\,\mathrm{cm}\) spacing. Walls are modeled as \(3\,\mathrm{m}\)-tall brick with default WinProp material settings. Furniture is manually drawn per scene, with heights randomly chosen from \(\{0.5, 1.0, 1.5, 2.0\}\,\mathrm{m}\) and fir-wood material. Each radio-map sample contains one omnidirectional transmitter with height in \([0, 3]\,\mathrm{m}\), while receiver maps are generated for heights from \(0.5\,\mathrm{m}\) to \(2.0\,\mathrm{m}\) in \(0.1\,\mathrm{m}\) steps, giving \(16\) horizontal slices per 3D map.

The dataset’s height embedding is defined by

$$
v_{\rm env}=\frac{h_{\rm env}+\varepsilon}{h_{\rm env,max}+\varepsilon},
$$

where \(h_{\rm env,max}=3\,\mathrm{m}\) and \(\varepsilon = 0.1\,\mathrm{m}\), with pixels having no object set to zero. Pathloss is generated with DPM according to

$$
P_{L,\rm dB}
= 20\log_{10}\!\Bigl(\tfrac{4\pi}{\lambda}\Bigr)
+10\,p\log_{10}(l)
+\sum_{i=1}^{N_f} f(\varphi_i,i)
+\sum_{j=1}^{N_t} t_j
-\Omega,
$$

where \(\lambda\) is wavelength, \(p\) is the path exponent, \(l\) is total path length, \(N_f\) is the number of directional interactions, \(f(\varphi_i,i)\) is the corresponding loss term, \(N_t\) is the number of penetrated walls, \(t_j\) is wall loss, and \(\Omega\) is the wave-guiding factor. The signal parameters are carrier \(5.9\,\mathrm{GHz}\), bandwidth \(W=10\,\mathrm{MHz}\), transmit power \(P_{\rm Tx}=23\,\mathrm{dBm}\), noise spectral density \(-174\,\mathrm{dBm/Hz}\), and receiver noise figure \(0\,\mathrm{dB}\), giving an approximate noise floor of \(-127\,\mathrm{dBm}\).

Sampling density is much higher than in the two measurement-based datasets. Each slice contains \(256 \times 256\) points, each 3D map contains \(16 \times 256 \times 256 = 1{,}048{,}576\) measurement points, and the full collection contains approximately \(3.36 \times 10^9\) pathloss samples. Data are organized into wall-layout images, furniture-layout images, transmitter-location images, and radio-map arrays of shape \(16 \times 256 \times 256\), together with metadata fields such as `building_id`, `furniture_id`, `tx_index`, `height_levels`, `pixel_spacing_x`, `pixel_spacing_y`, and `material_channels`.

The recommended workflow is to load wall, furniture, and transmitter images as three channels, stack them into a tensor of shape \((3,256,256)\), use the corresponding radio-map tensor \((16,256,256)\) as ground truth, split the \(3{,}200\) samples into \(80\%\) train, \(10\%\) validation, and \(10\%\) test at building level, optionally apply flips and rotations, and train a U-Net or R\(^2\)Net-In model. Evaluation uses NMSE, RMSE, SSIM, and PSNR. On the 3DiRM3200 test set, R\(^2\)Net-In reports NMSE \(=0.0268\), RMSE \(=0.0395\) in normalized units, SSIM \(=0.8908\), and PSNR \(=29.65\,\mathrm{dB}\), with throughput of approximately \(3\) maps per second on an Intel i5-11400F CPU and inference time of approximately \(46.8\,\mathrm{s}\) per \(\mathrm{km}^2\).

## 7. Comparative interpretation and recurring misconceptions

The most common misconception surrounding 3DiRM3200 is that it names a single standardized benchmark. The available arXiv record shows otherwise: the term has been attached to a sparse drone-supported beacon dataset [2109.06923], a real multimodal LiDAR-plus-RSSI indoor mapping dataset [2511.00494], and a dense fully simulated 3D pathloss benchmark [2605.18127]. These resources differ in sensing modality, spatial support, frequency, target variable, and evaluation protocol.

The multimodal 2025 dataset is centered on alignment between geometry and measured RSSI in a multi-room office-and-corridor environment, including occupancy effects. The 2021 drone dataset is centered on autonomous acquisition of beacon observations at discrete 3D waypoints in a compact apartment living-room volume. The 2026 simulated benchmark is centered on dense 3D pathloss estimation across floorplans using DPM-generated supervision and height embedding for deep networks. A plausible implication is that cross-paper comparison of reported errors is generally not direct: the underlying tasks range from RSSI interpolation in \(\mathrm{dBm}\) to normalized pathloss-field regression, and the datasets span real measurement noise, occupancy dynamics, and purely simulated propagation.

Taken together, the three usages map out complementary strands of indoor radio-map research. One strand emphasizes real geometry-aware multimodal sensing, another emphasizes autonomous data acquisition in three dimensions, and a third emphasizes large-scale supervised learning on dense volumetric radio maps. The shared label 3DiRM3200 therefore functions less as a unique dataset name than as a family name applied to technically different resources in indoor REM research.

Source: https://www.emergentmind.com/topics/3d-indoor-radio-map-dataset-3dirm3200