manojkarnekar/construction-traversability-dataset

Construction Traversability Dataset

A construction-site RGB semantic segmentation dataset developed for research on terrain understanding, traversability estimation, and multimodal RGB–LiDAR perception for mobile robots.

Overview

This dataset contains RGB images and pixel-wise semantic segmentation masks collected in construction-site environments. The dataset is intended to support research on construction-site scene understanding and traversability-aware robot perception.

The release also includes the camera/LiDAR calibration used for RGB–LiDAR projection and the fine-tuned semantic segmentation model used in the associated research.

Dataset Contents

construction-traversability-dataset/
├── README.md
├── LICENSE
├── dataset.yaml
│
├── images/
│   ├── train/
│   └── val/
│
├── masks/
│   ├── train/
│   └── val/
│
├── calibration/
│   ├── camera_intrinsics.yaml
│   └── lidar_camera_extrinsics.yaml
│
├── model/
│   └── model_config.yaml
│   └── yolo26_sem_construction.pt

RGB Images

RGB images are provided under:

• images/train/
• images/val/

Semantic Masks

Each RGB image has a corresponding pixel-wise semantic segmentation mask under the masks/ directory.

The mask uses integer class IDs, with the class definitions given below. The RGB image and mask use corresponding filenames.

This is a pixel-wise semantic segmentation representation rather than COCO polygon/RLE annotation format.

Semantic Classes

The dataset uses the following 28-class construction-site taxonomy:

ID	Class
0	animal
1	building
2	ceiling
3	concrete_blocks
4	construction_machinery
5	debris
6	fence
7	flat_road
8	material_pile
9	not_visible
10	person
11	pillar
12	pipes
13	pit
14	pole
15	ponding_concrete
16	puddle
17	rebar
18	rocky_terrain
19	scaffolding
20	sky
21	terrain
22	tiles
23	vegetation
24	vehicle
25	wall
26	wet_mud
27	wooden_planks

The class IDs correspond to the dataset taxonomy used by the semantic segmentation and RGB–LiDAR fusion pipeline.

Dataset Configuration

dataset.yaml contains the train/validation paths and class names used by the training pipeline.

Example:

path: .
train: images/train
val: images/val
names:
  0: animal
  1: building
  ...

The repository's dataset.yaml should be treated as the authoritative training configuration.

Camera and LiDAR Calibration

The calibration/ directory contains the sensor parameters used by the RGB–LiDAR projection pipeline.

Camera Intrinsics

calibration/camera_intrinsics.yaml contains the OAK-D RGB camera intrinsic matrix used for projection.

The projection code uses a 640 × 480 image size and:

fx = 513.8645629882812
fy = 513.7389526367188
cx = 316.9952392578125
cy = 247.48223876953125

The supplied calibration file does not specify distortion coefficients.

LiDAR–Camera Extrinsics

calibration/lidar_camera_extrinsics.yaml contains the transforms used to relate the Livox LiDAR and OAK-D RGB camera frames through the base_footprint frame, including the optical-frame correction used by the fusion pipeline.

These parameters document the transforms used by the released processing pipeline. They should not be interpreted as an independently certified metrology calibration unless otherwise stated.

Fine-Tuned Semantic Segmentation Model

The model/ directory contains the fine-tuned semantic segmentation model used for the construction-site taxonomy.

The model is based on the YOLO26 semantic segmentation architecture and is fine-tuned for the 28 construction-site classes listed above.

The model is provided to facilitate reproducibility of the semantic segmentation and RGB–LiDAR fusion experiments.

Raw ROS 2 Data

The complete ROS 2 recordings used during data collection may be released separately because of their large file size.

Raw ROS bag repository:
https://huggingface.co/datasets/manojkarnekar/construction-rosbags The raw recordings are intended to provide the original sensor data and timestamps needed for reproduction and further research.

Data Format

The released annotated data is organized as paired RGB images and pixel-wise semantic masks.

This format is intentionally kept simple so that it can be used directly by semantic segmentation pipelines without requiring conversion from COCO polygon/RLE annotations.

Researchers who require another annotation format may convert the masks to their preferred representation.

Intended Use

This dataset is released under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license. Commercial use is not permitted. See LICENSE for the full license terms.

Possible applications include:

• construction-site semantic segmentation
• terrain and traversability perception
• RGB–LiDAR fusion
• mobile robot navigation
• construction-site scene understanding
• semantic costmap generation
• multimodal robotic perception

License

The dataset and associated materials in this repository are released under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license.

Commercial use is not permitted under this license.

See LICENSE for the full license text.

License information: https://creativecommons.org/licenses/by-nc/4.0/

Disclaimer

The dataset is provided for research purposes. No guarantee is made regarding the completeness, accuracy, or suitability of the data for a particular application. Users are responsible for validating the data before deploying models or systems based on it in real-world environments.

Contact

manojkarnekar1@gmail.com