Datasets:

dynamic-maps
/

hard-intersection-multimodal-sample

	---
	license: cc-by-4.0
	pretty_name: Hard Intersection Multimodal Samples
	language:
	- en
	size_categories:
	- 1k<n<10k
	task_categories:
	- image-to-3d
	- image-classification
	- image-segmentation
	- depth-estimation
	- object-detection
	- other
	annotations_creators:
	- expert-generated
	- human-annotated
	tags:
	- autonomous-driving
	- multimodal
	- lidar
	- camera
	- gnss
	- trajectory
	- hdmap
	- lanelet2
	- opendrive
	- 3dgs
	- semantic-segmentation
	- point-cloud
	- geospatial
	- intersection
	- urban-driving
	- digital-twin
	- E2E
	- simulation
	---

	# Hard Intersection Multimodal Samples

	## Dataset Summary
	Hard Intersection Multimodal Samples is a curated multimodal dataset of accident-prone urban intersection in Japan for autonomous driving research and development.
	It provides multi-camera images, trajectory, HD maps, semantic annotations, point cloud data, and 3DGS assets.
	A comprehensive, high-precision multimodal dataset of public road environments captured via an industrial-grade MMS equipped with high-performance IMU, GNSS, LiDAR, and cameras.
	Features raw sensor data (images, point clouds, trajectories), processed HDMaps, metric 3DGS, and HDMap-derived semantic images/point clouds.

	---

	## Dataset Description
	### Overview
	This dataset focuses on challenging real-world urban intersection and is designed to support research in perception, mapping, and scene understanding.
	The current release include the following intersection:
	- Takanawadai, Tokyo, Japan
	- The Takanawadai intersection concentrates multiple adverse driving conditions into a single spot: sensor blind spots at the crest of a hill, an irregular six-way intersection with a sharp curve, and other vehicles crossing centerlines on narrow roads.
	- By creating a 3D model of this real-world, accident-prone location—which easily triggers errors even in human drivers—it serves as the ideal "safety benchmark" to test whether autonomous systems can successfully navigate extreme edge cases.
	- Even in the real world, this intersection historically ranked as the second worst in Tokyo for traffic accidents.

	<div style="display:flex; gap:0; margin:0; padding:0;">
	<img src="./assets/26047a_Record084_260217_055911366_Camera_4.jpg" alt="camera4" style="width:33.333%; display:block; margin:0; padding:0;">
	<img src="./assets/26047a_Record084_260217_055911366_Camera_0.jpg" alt="camera0" style="width:33.333%; display:block; margin:0; padding:0;">
	<img src="./assets/26047a_Record084_260217_055911366_Camera_1.jpg" alt="camera1" style="width:33.333%; display:block; margin:0; padding:0;">
	</div>

	- The dataset combines:
	- Six synchronized camera views
	- Two synchronized camera views
	- Trajectory data
	- HD maps (OpenDRIVE/RoadRunner/Lanelet2/Vissim)
	- Semantic annotations
	- point cloud data
	- 3DGS reconstruction assets
	---


	### Key Features
	- Focused on high-risk real-world intersection
	- Multi-domain support:
	- sensor domain
	- map domain
	- semantic projection domain
	- reconstruction domain
	- 360-degree coverage with six cameras
	- Developer-friendly formats: OpenDRIVE/RoadRunner/Lanelet2/Vissim
	- LAS point clouds

	---

	### Modalities
	#### Raw / Primary Data
	- Six synchronized camera images (Camera0–Camera5)
	- Two synchronized camera images (Camera_1,_+15deg/Camera_2,_+15deg)
	- Point cloud data
	- Trajectory data

	#### Derived assets
	- HD maps (OpenDRIVE/RoadRunner/Lanelet2/Vissim)
	- Semantic image labels
	- Semantic point clouds with labels
	- 3DGS reconstruction assets
	---

	### Camera Setup
	- Six synchronized cameras (Camera0–Camera5)
	- Full 360-degree coverage

	- Two synchronized cameras (`Camera_1,_+15deg/Camera_2,_+15deg`)

	- The file naming convention is `{folder}_{record_name}_{date}_{hhmmssmmm}_Camera_{number}`
	- Camera0 is the front view, Camera1 is the front-right view, Camera2 is the rear-right view, Camera3 is the rear-left view, Camera4 is the front-left view, and Camera5 is the top view.
	- Additionally, for the two synchronized cameras: `Camera_1,_+15deg` represents the front view, and `Camera_2,_+15deg` represents the rear-right view.
	- Although the filename includes “+15deg,” it does not mean that the camera is installed in that direction.

	- calibration/images.txt: extrinsic parameters (camera poses for each image)
	- calibration/cameras.txt: intrinsic parameters (camera model and intrinsics)
	- In cameras.txt, for the PINHOLE model, PARAMS[] corresponds to: fx, fy, cx, cy.

	---

	### Trajectory Data
	- The file naming convention is `{folder}_{record_name}_{date}`
	- The dataset includes recordings in the following order of 26047_Record004, 26047_Record050, 26047a_Record004, and 26047a_Record084.
	- The trajectory data is defined in EPSG:6677.

	---
	### HD Map Formats
	![hdmap](assets/lanelet2.png)
	The HD map is provided in the format shown below.
	- OpenDRIVE(Ver1.4/1.6/1.8)
	- RoadRunner HD Map
	- Lanelet2
	- Vissim

	---

	### Point Cloud Data
	- Provided in LAS format
	- Multi-run aggregated (not single-frame LiDAR)
	- The point cloud data (.las) is defined in EPSG:6677, with elevations given as orthometric heights (above sea level).

	- Use cases:
	- Map-aligned geometry
	- Semantic projection base
	- Reconstruction workflows
	---

	### Semantic Annotations
	- Semantic Image: COCO JSON format
	- The semantic information of the image data and the corresponding images to which the information is assigned are defined in the COCO JSON format.
	<div style="display:flex; gap:0; margin:0; padding:0;">
	<img src="./assets/semantic_image_camera4.png" alt="semantic_camera4" style="width:33.333%; display:block; margin:0; padding:0;">
	<img src="./assets/semantic_image_camera0.png" alt="semantic_camera0" style="width:33.333%; display:block; margin:0; padding:0;">
	<img src="./assets/semantic_image_camera1.png" alt="semantic_camera1" style="width:33.333%; display:block; margin:0; padding:0;">
	</div>

	- Semantic Point Cloud: point-wise (LiDAR)
	![semantic_pointcloud](assets/semantic_point_cloud_data.png)
	- The semantic information of the Semantic Point Cloud is stored in the UserData field of the LAS format, as specified in the table below.
	<table>
	<thead>
	<tr>
	<th>Item</th>
	<th>Code</th>
	</tr>
	</thead>
	<tbody>
	<tr><td>road_surface</td><td>11</td></tr>
	<tr><td>traffic_island</td><td>12</td></tr>
	<tr><td>solid_white_line</td><td>21</td></tr>
	<tr><td>dashed_white_line</td><td>22</td></tr>
	<tr><td>solid_yellow_line</td><td>23</td></tr>
	<tr><td>dashed_yellow_line</td><td>24</td></tr>
	<tr><td>double_line</td><td>25</td></tr>
	<tr><td>straight_arrow</td><td>31</td></tr>
	<tr><td>left_arrow</td><td>32</td></tr>
	<tr><td>right_arrow</td><td>33</td></tr>
	<tr><td>left_and_straight_arrow</td><td>34</td></tr>
	<tr><td>right_and_straight_arrow</td><td>35</td></tr>
	<tr><td>pedestrian_crossing</td><td>41</td></tr>
	<tr><td>stop_bar</td><td>42</td></tr>
	<tr><td>traffic_calming_strip</td><td>43</td></tr>
	<tr><td>bus</td><td>44</td></tr>
	<tr><td>vertical_two_traffic_light</td><td>51</td></tr>
	<tr><td>horizontal_three_traffic_light</td><td>52</td></tr>
	<tr><td>horizontal_four_traffic_light</td><td>53</td></tr>
	<tr><td>wrong_way</td><td>61</td></tr>
	<tr><td>interstate_route</td><td>62</td></tr>
	<tr><td>other_blue</td><td>63</td></tr>
	<tr><td>other</td><td>71</td></tr>
	</tbody>
	</table>

	- These annotations are assigned based on HD Map attributes, and objects not defined in the HD Map, such as vehicles, are not annotated.

	---
	### 3DGS and Reconstruction Assets
	- 3DGS assets are included as reconstruction-oriented scene representations.
	- These assets are intended for scene reconstruction, scene visualization, synthetic data preparation, simulation-oriented environment understanding, and map-aligned scene asset generation.
	- 3DGS assets are defined in the same coordinate system as the point cloud data.
	- Data Collection Constraints: Constructed clean, static data by completely filtering out dynamic objects from highly congested public roads with constant vehicle and pedestrian traffic.
	- Environmental Complexity: Accurately reproduced the geometry of complex, dynamic environments—where standard 3DGS generation typically fails—by integrating high-precision LiDAR point clouds captured for HDMap creation.
	- System Scale: Scaled beyond single objects to cover large-scale road networks through an integrated image and point cloud pipeline. Furthermore, when paired with our HDMap data, it functions as a "Metric 3DGS" capable of precise real-world physical dimensions and positioning.

	![3dgs](assets/3dgs.png)
	[Please refer to the following viewer for 3DGS assets.](https://huggingface.co/spaces/dynamic-maps/hard-intersection-3dgs-sample)

	---
	## Dataset Structure
	```text
	Dataset
	├─annotation
	│ ├─semantic_images
	│ └─semantic_pointcloud
	├─calibration
	├─images
	├─maps
	│ ├─lanelet2
	│ ├─opendrive
	│ ├─roadrunnerhdmap
	│ └─vissim
	│ └─images
	├─pointcloud
	├─reconstruction
	└─trajectory
	```
	---

	## Intended Uses
	- multi-camera perception research
	- map-aware perception
	- semantic segmentation
	- map projection workflows
	- difficult-scene localization
	- trajectory analysis
	- point cloud semantic research
	- panorama-based perception research
	- reconstruction and 3D scene asset research
	- synthetic data or simulation preparation workflows

	---

	## Social Impact of Dataset
	- It is expected that an increase in the volume of this dataset will contribute to the following: Improvement of autonomous driving technology.
	- Reduction in storage capacity through the elimination of duplicate datasets.


	## Limitations
	- Limited to 1 intersection (not large-scale coverage)
	- LAS is aggregated multi-run data
	- Representation differences across formats
	- Images have been processed to protect personal information: faces have been mosaicked, and license plates have been masked as much as possible, although the masking may not be complete. Users may contact us if additional anonymization is required.
	- The semantic images are not provided for all images. Users may contact us if additional annotation is required.

	---

	## Difficulty Tags
	- occlusion_heavy
	- dense_traffic
	- complex_lane_topology
	- multi_phase_signal
	- unprotected_turn

	---

	## Citation
	```bibtex
	@dataset{hard_intersection_multimodal_samples_2026,
	title={Hard Intersection Multimodal Samples},
	author={Dynamic Map Platform Co., Ltd.},
	year={2026},
	publisher={Hugging Face}
	}
	```

	## Acknowledgements / Attribution
	This dataset was generated using several open-source models and libraries. We gratefully acknowledge the contributions of the original authors.

	### Models
	- Grounding DINO
	Liu et al., IDEA-Research, ECCV 2024
	License: Apache License 2.0
	https://github.com/IDEA-Research/GroundingDINO

	- OneFormer
	Jain et al., SHI-Labs, CVPR 2023
	License: MIT License
	https://github.com/SHI-Labs/OneFormer

	- ViTMatte
	Yao et al., HUST Vision Lab, Information Fusion 2024
	License: Apache License 2.0
	https://github.com/hustvl/ViTMatte

	### Pretrained Models
	- Grounding DINO Base
	IDEA-Research
	License: Apache License 2.0
	https://huggingface.co/IDEA-Research/grounding-dino-base

	- OneFormer Cityscapes Swin-L
	SHI-Labs
	License: MIT License
	https://huggingface.co/shi-labs/oneformer_cityscapes_swin_large

	- ViTMatte Base (Distinctions-646)
	HUST Vision Lab
	License: Apache License 2.0
	https://huggingface.co/hustvl/vitmatte-base-distinctions-646

	### 3D Gaussian Splatting Libraries
	- gsplat
	Ye, Li et al., UC Berkeley / Nerfstudio, JMLR 2025
	License: Apache License 2.0
	https://github.com/nerfstudio-project/gsplat

	- Splatfacto-W
	Xu et al., UC Berkeley / ShanghaiTech, arXiv 2024
	License: Apache License 2.0
	https://github.com/KevinXu02/splatfacto-wyy

	---

	This dataset contains only derived data outputs generated using the above tools.
	No original model weights or source code are redistributed.

	All rights and licenses of the original works remain with their respective authors.

	### Redistribution Notice

	This repository distributes dataset artifacts only.
	It does NOT include:
	- source code of the above models
	- pretrained model weights
	- third-party libraries

	Users must obtain those components separately from their original sources and comply with their respective licenses.

	## Feedback/Contact
	- Feedback is optional, but very welcome.
	- Contact: opensource@dynamic-maps.co.jp

	Dynamic Map Platform Co., Ltd. (DMP) is a Japan-based provider of high-precision 3D geospatial data and HD maps for automotive and infrastructure applications.
	Established with support from the Japanese government and major automakers, DMP has built strong relationships with both industry and public sectors.
	[LinkedIn](https://www.linkedin.com/company/dynamic-map-platform/)

	In the automotive domain,
	DMP works with global OEMs including Toyota, Honda, General Motors, Nissan, and SUBARU supporting production vehicles and advanced driving systems.
	[PDF](https://contents.xj-storage.jp/xcontents/AS05208/e3df1e47/56c8/4631/ada2/48e7db830726/140120250611587375.pdf)