Datasets:
File size: 3,288 Bytes
a316c85 ae0129b a316c85 ae0129b | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 | ---
license: cc-by-4.0
task_categories:
- depth-estimation
- robotics
language:
- en
tags:
- mmWave-radar
- FMCW
- imaging
- depth-image
size_categories:
- 10K<n<100K
---
# TextRadar Dataset
TextRadar is a multimodal indoor RF sensing dataset introduced in the our paper **“3D Occupancy-Language Fields from mmWave Radar.”**
It contains synchronized high-resolution mmWave radar (TI MMWCAS-RF-EVM), LiDAR, and IMU measurements collected using a mobile robotic platform across **7 indoor buildings**.
The dataset supports research in RF imaging, depth estimation, robot localization, and multimodal perception.
## Dataset Details
- **~20k synchronized RF/LiDAR samples**
- **7 indoor buildings**
- **25 robot trajectories**
- Approximately **461 GB of processed data**
- **Curated by:** Haowen Lai, Zhiwei Zheng, and Mingmin Zhao, WAVES Lab, University of Pennsylvania
- **License:** CC BY 4.0
### Links
- **Paper:** *3D Occupancy-Language Fields from mmWave Radar*, arXiv 2026
## Dataset Structure
The dataset contains two main components.
### Raw Data
Raw recordings include:
- mmWave radar I/Q measurements
- synchronized LiDAR measurements
- synchronized IMU measurements
- timestamps
The raw data are intended for researchers working on radar signal processing, motion estimation, beamforming, and RF imaging.
### Processed Data
Processed data include:
- 3D RF heatmaps
- LiDAR-derived range ground truth
- glass masks
These data can be used directly for training and evaluating machine-learning models without rerunning the full radar-processing pipeline.
## Recommended Split
The original TextRadar experiments use **leave-one-building-out (LOBO)** evaluation.
For each experiment:
- 6 buildings are used for training
- 1 building is held out for testing
This split is recommended when evaluating generalization to unseen environments.
## Data Collection
Data were collected using a robotic platform equipped with:
- **TI MMWCAS-RF-EVM** mmWave FMCW radar
- **Ouster 64-beam LiDAR**
- IMU
- rotating radar assembly
The rotating radar forms a synthetic cylindrical array for panoramic 3D RF imaging.
LiDAR measurements provide geometric ground truth, while synchronized radar measurements are processed into 3D RF heatmaps.
## Annotations
Range and geometric ground truth are derived primarily from synchronized LiDAR measurements.
## Uses
TextRadar is intended for research in areas including:
- RF and mmWave sensing
- 3D RF imaging
- depth and range estimation
- surface-normal estimation
- robot localization
- RF–LiDAR multimodal learning
- cross-building generalization
## Limitations
The dataset was collected in **7 indoor buildings** and may not generalize directly to outdoor environments.
Results may also depend on the specific radar configuration, sensor mounting, and acquisition setup used in TextRadar.
LiDAR is used as geometric ground truth but can be unreliable on transparent surfaces such as glass. Glass masks are included to help identify these regions.
## Citation
If you use TextRadar, please cite:
```bibtex
@article{TextRadar,
title={3D Occupancy-Language Fields from mmWave Radar},
author={Lai, Haowen and Zheng, Zhiwei and Zhao, Mingmin},
journal={arXiv preprint arXiv:2610.37929},
year={2026}
}
``` |