Nexdata Launches Autonomous Driving Dataset with 2D, 3D, and 4D Annotations

Nexdata, one of Japan's largest AI training data providers, today launched a high-precision training dataset optimized for autonomous driving and ADAS development in Japan.

Background: Shortage of 'Driving Data' in Autonomous Driving Development
Currently, in the field of autonomous driving algorithm development, not only the 'quantity' but also the 'quality' and 'regional adaptability' of training data are key to performance improvement. In Japan in particular, unique traffic sign standards, complex urban environments, and strict traffic rules make it difficult for general-purpose datasets collected abroad to improve recognition accuracy in real environments. Furthermore, it is difficult to procure multi-modal data that cross-sectionally synchronizes 2D images, 3D point clouds, and time-series lane information, forcing development teams to spend significant resources on data collection and pre-processing.

Solution: 'Ready-to-Use' Dataset Collected from Real-World Data
Nexdata has developed an autonomous driving multi-modal dataset based on real driving data collected in actual road environments within Japan. This dataset provides practical training data that can be used immediately in development by a closed-loop system that manages everything from data collection to annotation, quality verification, and delivery.

Dataset Overview
Multi-modal annotation data collected by running actual vehicles on urban and coastal roads in Japan, synchronizing LiDAR point clouds, 6-view synchronized RGB cameras, RTK-GNSS/IMU, and CAN bus signals at the millisecond level.
Because it comprehensively includes high-precision 3D object tracking boxes, 4D lane recognition annotations, and 2D traffic sign detection data, it can be used for improving the learning efficiency of environment recognition models, accuracy verification of object tracking algorithms, HD map construction support, and demonstration experiments of ADAS functions.

Annotation Content
- 2D Traffic Signs: Covers all regulatory, instruction, and warning signs used in Japan. Supports bilingual labels in Japanese and English (e.g., Road Closed / 通行止め), and comes standard with high-precision 2D bounding boxes, class attributes, and status flags (compatible with nighttime, damage, and occlusion). Images are provided in 4K resolution (4096×2160), and annotations are provided in JSON format.
- 3D Bbox Point Clouds: Recorded with 6 4K in-vehicle cameras, a 16-line LiDAR, and IMU/GNSS, synchronized in space-time (at the microsecond level). Includes 3D bounding box annotations for vehicles, pedestrians, two-wheelers, and obstacles. Includes external/internal parameter calibration files (calib.yaml), allowing immediate start of sensor fusion model training.
- 4D Lane Recognition: Provides 'time-series point cloud maps' that superimpose multiple LiDAR frames with high precision, rather than a single frame. Clearly reproduces the shapes of white and yellow lines without gaps, improving the learning efficiency of lane detection algorithms utilizing Intensity information. Also supports the generation of vector data for HD Maps.

Usage Scenarios
- Autonomous Driving Algorithm Development (Object detection, lane recognition, sign understanding)
- ADAS Function Verification (Test data for evaluating AEB, LKA, ACC, etc.)
- Simulation Environment Construction (Virtual space reproduction based on real-world data)

Customization Available
This dataset can be flexibly customized according to your specific development requirements and verification scenarios, including collection area, weather conditions, annotation items, and data formats, through our made-to-order service.