Voxel51/hard-intersection-multimodal-sample

Dataset Card for Hard Intersection Multimodal Sample

Dataset Details

Dataset Description

Hard Intersection Multimodal Sample is a curated multimodal dataset of an accident-prone six-way urban intersection in Tokyo, Japan (Takanawadai) captured with an industrial mobile mapping system. The dataset provides synchronized multi-camera views, LiDAR point clouds, vehicle trajectories, HD maps in multiple formats, and semantic annotations for autonomous driving research.

The Takanawadai intersection historically ranked as the second worst in Tokyo for traffic accidents, featuring multiple challenging conditions: sensor blind spots at a hill crest, an irregular six-way intersection with sharp curves, narrow roads where vehicles cross centerlines, and complex traffic signal phasing.

• Curated by: Dynamic Map Platform Co., Ltd.
• Funded by: Dynamic Map Platform Co., Ltd. (with support from the Japanese government and major automakers)
• Shared by: Dynamic Map Platform Co., Ltd.
• Language(s): English (metadata and documentation)
• License: CC-BY-4.0

Dataset Sources

• Repository: https://huggingface.co/datasets/dynamic-maps/hard-intersection-multimodal-sample
• Paper: No associated research paper (industry dataset)
• Demo: https://huggingface.co/spaces/dynamic-maps/hard-intersection-3dgs-sample (3DGS viewer)

Uses

Direct Use

This dataset is designed for:

• Multi-camera perception algorithm development and testing in challenging urban scenarios
• Map-aware perception systems that integrate HD maps with sensor data
• Semantic segmentation model training and evaluation on road infrastructure
• Localization and trajectory analysis in difficult urban environments
• Point cloud semantic understanding research
• 360-degree panoramic perception system development
• 3D scene reconstruction and synthetic data generation workflows
• HD map projection and validation against real sensor observations
• Autonomous vehicle safety benchmarking in high-risk intersection scenarios

The synchronized multi-camera setup with calibration data enables stereo and multi-view geometry research. The co-temporal trajectory and LIDAR data support sensor fusion work.

Out-of-Scope Use

• Large-scale geographic coverage studies (limited to 1 intersection)
• Real-time LIDAR frame-by-frame processing (point cloud is multi-run aggregated, not per-episode temporal scans)
• Cross-format HD map consistency research without custom projection code (maps use incompatible local coordinate systems)
• Privacy-sensitive applications without additional review (faces and license plates have been anonymized but masking may not be complete)

Dataset Structure

This is a grouped FiftyOne dataset with 4 groups (one per driving episode through the intersection). The media_type is group and the default group slice is Camera_0.

Dataset Topology

Each group represents a single ~12-25 second driving pass through the intersection, captured on February 17, 2026 (inferred from 260217 in filenames). The 4 episodes are:

• 26047_Record004_260217 (28 frames, 13.7s)
• 26047_Record050_260217 (26 frames, 12.6s)
• 26047a_Record004_260217 (30 frames, 19.4s)
• 26047a_Record084_260217 (28 frames, 11.9s)

Group slices: Camera_0, Camera_1, Camera_2, Camera_3, Camera_4, Camera_5 (videos), point_cloud (3D)

Total samples: 28 (24 video samples + 4 3D point cloud samples)

Sample Fields

Field	FiftyOne type	Description
filepath	StringField	Path to video (.mp4) or FO3D scene file (.fo3d)
group	Group	Group membership linking all slices of an episode
episode_id	StringField	Episode identifier (e.g. 26047_Record004_260217)
camera	StringField	Camera identifier (e.g. Camera_0) for video slices
metadata	VideoMetadata or SceneMetadata	Video properties (duration, FPS, resolution, encoding) or 3D scene properties (asset counts, size)

Frame Fields (Video Samples Only)

Field	FiftyOne type	Description
ground_truth	Detections	COCO semantic annotations (29 classes: road markings, traffic signs, traffic lights, etc.)
hd_map	Polylines	Lanelet2 HD map elements projected from 3D world coordinates to camera view (Camera_0-4)
trajectory	Polylines	Vehicle GNSS/IMU trajectory projected from 3D world coordinates to camera view (Camera_0 only)

Label Types and Mapping Decisions

1. COCO Semantic Annotations → fo.Detections

The source provides COCO JSON with bounding boxes and polygon segmentations for 29 road infrastructure classes (lane lines, arrows, traffic lights, signs, etc.). Mapped to fo.Detections as frame-level labels on videos.

• Coverage: Only cameras 0-4; Camera 5 (top view) is never annotated. Only 2 of 4 episodes have annotations (26047_Record004, 26047_Record050 — 270 of 896 images, ~30%).
• Conversion: COCO bbox absolute pixels [x,y,w,h] → FiftyOne relative [x/w, y/h, w/w, h/h]. COCO category_id (1-indexed) mapped to class names. Polygon segmentation preserved in detection objects.

2. Lanelet2 HD Map → fo.Polylines

HD map polylines (lane boundaries, center lines, stop lines, crosswalks) from Lanelet2 OSM format projected into each camera view using COLMAP calibration. Mapped to fo.Polylines as frame-level labels.

• Projection: Lanelet2 nodes carry both lat/lon (WGS84) and local Transverse Mercator coordinates. These are distinct from the point cloud's EPSG:6677 coordinate system. The ingestion reprojects via the known origin (lat 35.632°, lon 139.730°), transforms to camera frame using COLMAP world-to-camera extrinsics (quaternion + translation), then projects to 2D image coordinates. Polyline segments are clipped to image bounds and split into contiguous runs where points remain visible.
• Coverage: All frames, all cameras 0-4. Camera 5 projections were skipped (top-down view geometry).

3. Vehicle Trajectory → fo.Polylines (Front Camera Only)

The vehicle's pose stream (GNSS/IMU, 200 Hz, EPSG:6677 coordinates: Easting, Northing, Height) is projected onto the front camera (Camera_0) only, using the identical world-to-camera projection and image-bounds clipping logic as the HD map (reuses the same COLMAP extrinsics/intrinsics per frame). Stored as a frame-level fo.Polylines field, trajectory, showing the vehicle's full driving path through the intersection as seen from the windshield view.

• Coverage: 100% of Camera_0 frames across all 4 episodes. Not projected onto Camera_1-5 (README identifies Camera_0 as the front view; the other 5 cameras face other directions where the trajectory overlay is less meaningful).
• An earlier iteration visualized the trajectory as colored spheres inside the 3D point cloud scene, but this was dropped in favor of the simpler 2D polyline-on-front-camera representation.

4. Point Cloud → FO3D Scenes

The aggregated LiDAR point cloud (EPSG:6677, 35M points) was spatially subsetted to points within 30m of each episode's trajectory using a KDTree distance filter, converted from LAS to PCD format (with RGB color and intensity), then wrapped in FO3D scene files containing only the point cloud asset. Large point clouds (>10M points) were downsampled to 10M for app performance.

• Color: Source LAS has no real RGB variation (all white). Height-based coloring is the default shading mode in FO3D.

Video Encoding

Videos were created from JPEG image sequences using ffmpeg with H.264 encoding:

• Frame rate: Calculated from actual image timestamps (1.49 - 2.26 FPS per episode, not a fixed rate)
• Resolution: 2048×2464 (portrait orientation, matching source images)
• Codec: H.264/avc1, CRF 23, YUV420p pixel format

Dataset-Level Metadata (dataset.info)

• crs: "EPSG:6677" (Japan Plane Rectangular CS IX)
• episodes: List of 4 episode IDs
• notes: Description of dataset structure, video encoding, label types, and point cloud subsetting

Parsing Decisions

1. Grouped dataset vs. multimodal (MCAP): Chose grouped dataset over MCAP-based multimodal because: (a) labels (COCO detections, HD map polylines) integrate natively as FiftyOne Label fields on video frames; (b) timeline synchronization was ambiguous (trajectory CSV Time[s] doesn't trivially align with image filename timestamps); (c) FO3D 3D scenes work directly in groups, whereas MCAP 3D tiles have no native detection overlay.

2. Video vs. image samples: Encoded image sequences as H.264 videos (one per camera per episode) to match the temporal nature of the data and reduce sample count (4 episodes × 6 cameras = 24 video samples vs. 672 image samples).

3. Point cloud subsetting: The full scene point cloud (35M points, 1.2 GB LAS) is shared across all episodes. To make it episode-relevant and manageable, subsetted to points within 30m of each episode's vehicle trajectory (yields 10-32M points, 150-500 MB PCD per episode).

4. Auxiliary cameras excluded: The Camera_1,_+15deg and Camera_2,_+15deg image pairs (~224 images) were excluded because they are not time-synchronized with the main 6-camera rig (independent timestamps, occasional single-camera-only frames) and have zero semantic annotations.

5. Semantic point cloud not imported: The semantic_pointcloud/*.las file (per-point class codes in LAS user_data field) exists once for the scene, not per-episode. FiftyOne's 3D visualizer does not have native semantic point cloud labels analogous to 2D segmentation masks; this was left as a dataset-level auxiliary asset rather than forced into FO3D scenes.

6. HD map coordinate reprojection: Lanelet2 OSM nodes use a local Transverse Mercator frame (85000/43900) that is numerically incompatible with the trajectory/point cloud EPSG:6677 frame (-9300/-40800). The ingestion computes the geodetic origin from OpenDRIVE's geoReference tag (lat 35.632°, lon 139.730°), then applies the inverse T-Mercator projection (pyproj) to obtain lat/lon, then forward-projects to EPSG:6677 to align with camera extrinsics.

7. COLMAP extrinsics convention: calibration/images.txt provides world-to-camera transforms (quaternion, translation). These were used directly for projection — FiftyOne does not store camera poses, only the projected 2D labels.

8. Class remapping: COCO category_id values (1-indexed) were mapped to string class names using the embedded categories list in the COCO JSON. FiftyOne stores class names, not integer IDs.

9. Trajectory as 2D polyline, not 3D scene geometry: The vehicle trajectory was projected onto the front camera (Camera_0) as a fo.Polylines field rather than rendered as geometry inside the FO3D point cloud scene. This keeps the trajectory as an inspectable, per-frame FiftyOne Label (filterable/visualizable in the App's label sidebar) and reuses the existing HD map camera-projection pipeline, rather than adding scene-graph objects to the 3D asset.

Classes (29 Total)

road_surface, lane_line, Pavement Striping, intersection, guardrail_line, solid_white_line, dashed_white_line, solid_yellow_line, dashed_yellow_line, double_line, Straight Arrow, left arrow, right arrow, Pedestrian Crossing, Stop Bar, Deceleration Marking, Horizontal Three Traffic Light, Other – Blue – Information, warning_sign, Interstate Route, back_of_sign, Left and Straight Arrow, Right and Straight Arrow, 4 Horizontal Traffic Light, Other, bus, Vertical Two Traffic Light, Wrong Way, Parallel Parking.

Classes are derived from HD map attributes; dynamic objects (vehicles, pedestrians) are not annotated.

Dataset Creation

Curation Rationale

The Takanawadai intersection in Tokyo was selected because it concentrates multiple adverse driving conditions that challenge both human drivers and autonomous systems:

• Sensor blind spots at the crest of a hill
• Irregular six-way intersection with sharp curves rather than standard four-way perpendicular crossings
• Narrow roads where vehicles frequently cross centerlines
• Complex traffic signal phasing with multiple unprotected turn scenarios
• Dense traffic and heavy occlusion from surrounding buildings and terrain

Historically, this intersection ranked second worst in Tokyo for traffic accidents. It serves as an ideal "safety benchmark" location to test whether autonomous systems can handle extreme edge cases that frequently cause errors even for experienced human drivers.

Source Data

Data Collection and Processing

Data was captured using an industrial-grade mobile mapping system (MMS) equipped with:

• 6 synchronized cameras (2048×2464, RGB, 8-bit JPEG) providing 360° coverage
• High-precision LiDAR scanner producing multi-return point cloud data
• IMU (Inertial Measurement Unit) and GNSS (Global Navigation Satellite System) for trajectory recording at 200 Hz
• COLMAP-based structure-from-motion calibration providing per-image extrinsics (camera poses) and intrinsics (focal length, principal point)

Coordinate system: EPSG:6677 (Japan Plane Rectangular Coordinate System IX), a projected CRS optimized for Tokyo-area surveys with elevations as orthometric heights (above sea level).

The point cloud is multi-run aggregated — combined from multiple driving passes to achieve complete scene coverage — rather than single-frame LiDAR. The 3D Gaussian Splatting (3DGS) reconstruction was generated by integrating the LiDAR geometry with camera imagery, filtering out dynamic objects to produce a clean, static scene model.

Data collection constraint: All dynamic objects (vehicles, pedestrians) were filtered from the 3DGS and semantic point cloud to maintain a static baseline. This was necessary due to the highly congested public road environment with constant traffic.

Who are the source data producers?

Dynamic Map Platform Co., Ltd. (DMP), a Japan-based provider of high-precision 3D geospatial data and HD maps for automotive and infrastructure applications. DMP was established with support from the Japanese government and major automakers (Toyota, Honda, General Motors, Nissan, SUBARU) and has built relationships with both industry and public sectors. The company specializes in meter-level map accuracy for advanced driver assistance systems (ADAS) and autonomous vehicle development.

Annotations

Annotation process

Semantic image annotations (COCO format, 29 classes) were generated using the following open-source models, with labels derived from HD map attributes:

1. Grounding DINO Base (IDEA-Research, ECCV 2024, Apache-2.0): Zero-shot object detection to localize map elements in images
2. OneFormer Cityscapes Swin-L (SHI-Labs, CVPR 2023, MIT): Universal image segmentation to produce polygon masks
3. ViTMatte Base (HUST Vision Lab, Information Fusion 2024, Apache-2.0): High-quality matting for refining boundaries

Semantic point cloud annotations (23 classes in LAS user_data field) were assigned based on HD map attributes projected into 3D. Objects not defined in the HD map (e.g., vehicles, pedestrians, poles) were not annotated.

3DGS reconstruction was generated using:

• gsplat (UC Berkeley / Nerfstudio, JMLR 2025, Apache-2.0): 3D Gaussian splatting library
• Splatfacto-W (UC Berkeley / ShanghaiTech, arXiv 2024, Apache-2.0): Robust 3DGS for large-scale outdoor scenes

No human annotators were involved in the production of semantic labels; all annotations are model-generated outputs aligned to HD map geometry.

HD maps (OpenDRIVE v1.4/1.6/1.8, Lanelet2, RoadRunner, Vissim, Unreal Engine) were created through DMP's proprietary mapping pipeline using the captured MMS data. Map formats represent the same physical intersection in different industry-standard schemas.

Who are the annotators?

All semantic annotations were generated algorithmically using the pretrained models listed above, without manual human annotation. DMP's engineering team operated the MMS data collection vehicle and produced the HD maps.

Personal and Sensitive Information

Images have been processed to protect personal information:

• Faces: Mosaicked (blurred/pixelated)
• License plates: Masked where detected

Note: Anonymization may not be complete. Users requiring additional privacy protection should contact Dynamic Map Platform Co., Ltd. at opensource@dynamic-maps.co.jp.

Citation

BibTeX:

@dataset{hard_intersection_multimodal_samples_2026,
  title={Hard Intersection Multimodal Sample},
  author={Dynamic Map Platform Co., Ltd.},
  year={2026},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/datasets/dynamic-maps/hard-intersection-multimodal-sample}}
}

APA:

Dynamic Map Platform Co., Ltd. (2026). Hard Intersection Multimodal Sample [Data set]. Hugging Face. https://huggingface.co/datasets/dynamic-maps/hard-intersection-multimodal-sample

More Information

Difficulty tags (dataset-level characteristics):

• occlusion_heavy: Significant visual occlusion from buildings, terrain, and traffic
• dense_traffic: High vehicle and pedestrian density
• complex_lane_topology: Irregular six-way intersection with sharp curves
• multi_phase_signal: Complex traffic signal timing with multiple phases
• unprotected_turn: Turn movements without dedicated signal protection

Limitations:

• Limited to 1 intersection (not large-scale geographic coverage)
• LiDAR point cloud is multi-run aggregated, not per-episode temporal scans
• Semantic image annotations cover only 30% of images (270 of 896, cameras 0-4 only, 2 of 4 episodes)
• Camera 5 (top view) and auxiliary +15deg cameras have no semantic annotations
• HD map formats use different coordinate representations; cross-format workflows require custom projection code
• Anonymization (face/plate masking) may not be 100% complete

Related resources:

• 3DGS viewer: https://huggingface.co/spaces/dynamic-maps/hard-intersection-3dgs-sample
• Synthetic trajectory videos: Straight path, Right-turn path

Dataset Card Authors

Harpreet Sahota

Dataset Card Contact

opensource@dynamic-maps.co.jp