Datasets:

ArXiv:
semantic2d / dataset /README.md
zzuxzt's picture
Upload semantic2d dataset
d46530b verified
# Semantic2D Dataset
A 2D lidar semantic segmentation dataset for mobile robotics applications. This is the first publicly available 2D lidar semantic segmentation dataset, featuring point-wise annotations for nine indoor object categories across twelve distinct environments.
**Associated Paper:** *Semantic2D: Enabling Semantic Scene Understanding with 2D Lidar Alone*
**Authors:** Zhanteng Xie, Yipeng Pan, Yinqiang Zhang, Jia Pan, Philip Dames
**Institutions:** The University of Hong Kong, Temple University
**Video:** https://youtu.be/P1Hsvj6WUSY
**GitHub:** https://github.com/TempleRAIL/semantic2d
---
## Dataset Overview
| Property | Value |
|----------|-------|
| Total Data Tuples | ~188,007 |
| Recording Rate | 20 Hz |
| Total Duration | ~131 minutes |
| Environments | 12 indoor environments |
| Buildings | 7 buildings (4 at Temple, 3 at HKU) |
| LiDAR Sensors | 3 types |
| Semantic Classes | 9 + background |
### Semantic Classes
| ID | Class | Description |
|----|-------|-------------|
| 0 | Background/Other | Unclassified objects |
| 1 | Chair | Chairs and seating |
| 2 | Door | Doors and doorways |
| 3 | Elevator | Elevator doors |
| 4 | Person | Dynamic pedestrians |
| 5 | Pillar | Structural pillars |
| 6 | Sofa | Sofas and couches |
| 7 | Table | Tables |
| 8 | Trash bin | Trash cans |
| 9 | Wall | Walls and partitions |
### LiDAR Sensors
| Sensor | Robot | Location | Range (m) | FOV (deg) | Angular Res. (deg) | Points |
|--------|-------|----------|-----------|-----------|-------------------|--------|
| Hokuyo UTM-30LX-EW | Jackal | Temple | [0.1, 60] | 270 | 0.25 | 1,081 |
| WLR-716 | Customized | HKU | [0.15, 25] | 270 | 0.33 | 811 |
| RPLIDAR-S2 | Customized | HKU | [0.2, 30] | 360 | 0.18 | 1,972 |
---
## Directory Structure
```
semantic2d_dataset_2025/
├── README.md
├── rosbags/ # Original ROS bag files, only used for reference or to extract other data, such as RGB, depth, goal, etc.
│ ├── 2024-04-04-12-16-41.bag # Temple Engineering Lobby
│ ├── 2024-04-04-13-48-45.bag # Temple Engineering 6th floor
│ ├── 2024-04-04-14-18-32.bag # Temple Engineering 9th floor
│ ├── 2024-04-04-14-50-01.bag # Temple Engineering 8th floor
│ ├── 2024-04-11-14-37-14.bag # Temple Engineering 4th floor
│ ├── 2024-04-11-15-24-29.bag # Temple Engineering Corridor
│ ├── 2025-07-08-13-32-08.bag # Temple SERC Lobby
│ ├── 2025-07-08-14-22-44.bag # Temple Gladfelter Lobby
│ ├── 2025-07-18-17-43-11.bag # Temple Mazur Lobby
│ ├── 2025-11-10-15-53-51.bag # HKU Chow Yei Ching 4th floor
│ ├── 2025-11-11-21-27-17.bag # HKU Centennial Campus Lobby
│ └── 2025-11-18-22-13-37.bag # HKU Jockey Club 3rd floor
└── semantic2d_data/ # Semantic2D Segmentation Data files
├── dataset.txt # dataset index for each folder
├── 2024-04-04-12-16-41/ # Temple Engineering Lobby # Temple environments (Hokuyo)
├── 2024-04-04-13-48-45/ # Temple Engineering 6th floor
├── 2024-04-04-14-18-32/ # Temple Engineering 9th floor
├── 2024-04-04-14-50-01/ # Temple Engineering 8th floor
├── 2024-04-11-14-37-14/ # Temple Engineering 4th floor
├── 2024-04-11-15-24-29/ # Temple Engineering Corridor
├── 2025-07-08-13-32-08/ # Temple SERC Lobby
├── 2025-07-08-14-22-44/ # Temple Gladfelter Lobby
├── 2025-07-18-17-43-11/ # Temple Mazur Lobby
├── 2025-11-10-15-53-51/ # HKU Chow Yei Ching 4th floor # HKU environments (WLR-716 + RPLIDAR)
├── 2025-11-11-21-27-17/ # HKU Centennial Campus Lobby
└── 2025-11-18-22-13-37/ # HKU Jockey Club 3rd floo
```
---
## Semantic2D Segmentation Data Folder Structure
### Temple University Environments (Hokuyo UTM-30LX-EW)
Folders: `2024-04-04-*` and `2024-04-11-*` (6 environments)
```
2024-04-04-12-16-41/ # Engineering Lobby
├── train.txt # Training split filenames (70%)
├── dev.txt # Validation split filenames (10%)
├── test.txt # Test split filenames (20%)
├── scans_lidar/ # LiDAR range data
│ └── *.npy # Shape: (1081,) - range values in meters
├── intensities_lidar/ # LiDAR intensity data
│ └── *.npy # Shape: (1081,) - intensity values
├── line_segments/ # Extracted line features
│ └── *.npy # Line segments [x1,y1,x2,y2] per line
├── positions/ # Robot poses
│ └── *.npy # Shape: (3,) - [x, y, yaw] in map frame
├── velocities/ # Velocity commands
│ └── *.npy # Shape: (2,) - [linear_x, angular_z]
├── semantic_label/ # Point-wise semantic labels
│ └── *.npy # Shape: (1081,) - class ID per point
├── semantic_scan/ # Semantic scan visualization data
│ └── *.npy
├── final_goals_local/ # Navigation final goals
│ └── *.npy # Goal positions in local frame
├── sub_goals_local/ # Navigation sub-goals
│ └── *.npy # Sub-goal waypoints
└── 202404041210_eng_lobby_map/ # Environment map
├── 202404041210_eng_lobby.pgm # Occupancy grid map (PGM format)
├── 202404041210_eng_lobby.yaml # Map configuration
└── map_labelme/ # Semantic labeled map
├── img.png # Original map image
├── label.png # Semantic label image
├── label_viz.png # Colored visualization
├── label_names.txt # Class name list
└── map_labelme.json # LabelMe annotation file
```
### HKU Environments (WLR-716 + RPLIDAR-S2)
Folders: `2025-11-*` (3 environments)
```
2025-11-10-15-53-51/ # Chow Yei Ching 4th floor
├── train.txt # Training split (70%)
├── dev.txt # Validation split (10%)
├── test.txt # Test split (20%)
├── # WLR-716 LiDAR data (811 points, 270 deg FOV)
├── scans_lidar_wlr716/ # Range data
│ └── *.npy # Shape: (811,)
├── intensities_lidar_wlr716/ # Intensity data
│ └── *.npy # Shape: (811,)
├── line_segments_wlr716/ # Line segments
│ └── *.npy
├── semantic_label_wlr716/ # Semantic labels
│ └── *.npy # Shape: (811,)
├── # RPLIDAR-S2 data (1972 points, 360 deg FOV)
├── scans_lidar_rplidar/ # Range data
│ └── *.npy # Shape: (1972,)
├── intensities_lidar_rplidar/ # Intensity data
│ └── *.npy # Shape: (1972,)
├── line_segments_rplidar/ # Line segments
│ └── *.npy
├── semantic_label_rplidar/ # Semantic labels
│ └── *.npy # Shape: (1972,)
├── # Shared data (same for both sensors)
├── positions/ # Robot poses [x, y, yaw]
│ └── *.npy # Shape: (3,)
├── velocities/ # Velocity commands [vx, wz]
│ └── *.npy # Shape: (2,)
└── 202511101415_cyc_4th_map/ # Environment map
├── *.pgm # Occupancy grid
├── *.yaml # Map configuration
└── map_labelme/ # Semantic labels
└── ...
```
---
## Data Format Details
### LiDAR Scan Data (`.npy`)
```python
import numpy as np
# Load range data
scan = np.load('scans_lidar/0000001.npy') # Shape: (N,) where N = num_points
# Hokuyo: N=1081, angle_min=-135°, angle_max=135°
# WLR-716: N=811, angle_min=-135°, angle_max=135°
# RPLIDAR-S2: N=1972, angle_min=-180°, angle_max=180°
# Convert to Cartesian coordinates
angles = np.linspace(angle_min, angle_max, num=N)
x = scan * np.cos(angles)
y = scan * np.sin(angles)
```
### Semantic Labels (`.npy`)
```python
# Load semantic labels
labels = np.load('semantic_label/0000001.npy') # Shape: (N,)
# Each value is a class ID (0-9):
# 0: Background, 1: Chair, 2: Door, 3: Elevator, 4: Person,
# 5: Pillar, 6: Sofa, 7: Table, 8: Trash bin, 9: Wall
```
### Robot Pose (`.npy`)
```python
# Load robot pose
pose = np.load('positions/0000001.npy') # Shape: (3,)
x, y, yaw = pose[0], pose[1], pose[2] # Position and orientation in map frame
```
### Dataset Splits (`train.txt`, `dev.txt`, `test.txt`)
```
# Each line contains a .npy filename
0001680.npy
0007568.npy
0009269.npy
...
```
---
## Semantic2D labeling:
### 1. Data Collection (`dataset_collection.py`)
ROS node for collecting data from the robot during teleoperation.
**Key Features:**
- Subscribes to LiDAR scans (`/scan, /wj716_base/scan`, `/rplidar_base/scan`)
- Records range, intensity, line segments, positions, velocities
- Saves data at 20 Hz as `.npy` files
### 2. Semi-Automated Labeling Framework (SALSA)
**Algorithm:**
1. Load pre-labeled semantic map (from LabelMe)
2. For each LiDAR scan:
- Extract line features for robust alignment
- Apply ICP to refine robot pose
- Project LiDAR points to map frame
- Match points to semantic labels via pixel lookup
- Points in free space labeled as "Person" (dynamic objects)
**Configuration (modify in script):**
```python
DATASET_ODIR = "/path/to/raw/data"
MAP_ORIGIN = np.array([-82.0, -71.6, 0.0]) # From map YAML
MAP_RESOLUTION = 0.025
POINTS = 811 # Number of LiDAR points
AGNLE_MIN = -2.356 # Min angle (radians)
AGNLE_MAX = 2.356 # Max angle (radians)
```
---
## Environments Summary
| Environment | Location | Folder | LiDAR |
|-------------|----------|--------|-------|
| Engineering Lobby | Temple | `2024-04-04-12-16-41` | Hokuyo |
| Engineering 6th Floor | Temple | `2024-04-04-13-48-45` | Hokuyo |
| Engineering 9th Floor | Temple | `2024-04-04-14-18-32` | Hokuyo |
| Engineering 8th Floor | Temple | `2024-04-04-14-50-01` | Hokuyo |
| Engineering 4th Floor | Temple | `2024-04-11-14-37-14` | Hokuyo |
| Engineering Corridor | Temple | `2024-04-11-15-24-29` | Hokuyo |
| SERC Lobby | Temple | `2025-07-08-13-32-08` | Hokuyo |
| Gladfelter Lobby | Temple | `2025-07-08-14-22-44` | Hokuyo |
| Mazur Lobby | Temple | `2025-07-18-17-43-11` | Hokuyo |
| Chow Yei Ching 4th Floor | HKU | `2025-11-10-15-53-51` | WLR-716/RPLIDAR |
| Centennial Campus Lobby | HKU | `2025-11-11-21-27-17` | WLR-716/RPLIDAR |
| Jockey Club 3rd Floor | HKU | `2025-11-18-22-13-37` | WLR-716/RPLIDAR |
---
## ROS Bag Contents
The original ROS bags in `rosbags/` contain:
| Topic | Message Type | Description |
|-------|--------------|-------------|
| `/scan` or `/*/scan` | `sensor_msgs/LaserScan` | LiDAR scans |
| `/robot_pose` | `geometry_msgs/PoseStamped` | Robot pose |
| `/cmd_vel` | `geometry_msgs/Twist` | Velocity commands |
| `/tf` | `tf2_msgs/TFMessage` | Transforms |
| `/map` | `nav_msgs/OccupancyGrid` | Occupancy map |
| `/camera/*` | `sensor_msgs/Image` | RGB/Depth images |
| `/odom` | `nav_msgs/Odometry` | Odometry |
---
## Related Resources
- **SALSA (Dataset and Labeling Framework):** https://github.com/TempleRAIL/semantic2d
- **S3-Net (Segmentation Algorithm):** https://github.com/TempleRAIL/s3_net
- **Semantic CNN Navigation:** https://github.com/TempleRAIL/semantic_cnn_nav
- **Dataset Zenodo:** DOI: 10.5281/zenodo.18350696
---
## Citation
```bibtex
@article{xie2026semantic2d,
title={Semantic2D: Enabling Semantic Scene Understanding with 2D Lidar Alone},
author={Xie, Zhanteng and Pan, Yipeng and Zhang, Yinqiang and Pan, Jia and Dames, Philip},
journal={arXiv preprint arXiv:2409.09899},
year={2026}
}
```
---
## License
Please refer to the associated paper and GitHub repository for licensing information.
## Contact
- Zhanteng Xie: zhanteng@hku.hk