Toward Generating Realistic 3D Semantic Training Data for Autonomous Driving.

Area of Science:

• Robotics and Computer Vision
• Artificial Intelligence

Background:

• 3D semantic segmentation is vital for autonomous driving, but data annotation is a major challenge.
• Synthetic data generation methods face a domain gap and often use complex multi-resolution approaches.
• Diffusion models show promise for realistic data synthesis but have limitations in 3D applications.

Purpose of the Study:

• To develop a novel approach for generating high-quality, scene-scale 3D semantic data.
• To eliminate reliance on image projection or decoupled multi-resolution models in synthetic data generation.
• To evaluate the effectiveness of the generated synthetic data for training semantic segmentation models.

Main Methods:

• A novel generative approach for 3D semantic scene-scale data synthesis without image projection.
• Avoidance of decoupled, multi-resolution trained models, ensuring end-to-end generation.
• Thorough evaluation of synthetic data's utility in training semantic segmentation networks.

Main Results:

• Achieved more realistic 3D semantic scene data generation compared to state-of-the-art methods.
• Demonstrated improved semantic segmentation model performance when trained with the generated synthetic data alongside real labels.
• Showcased the potential of synthetic point clouds to augment existing datasets and reduce annotation costs.

Conclusions:

• The proposed method effectively generates realistic 3D semantic scene data, bridging the domain gap.
• Synthetic data generated by this method enhances the performance of semantic segmentation models.
• This approach offers a viable solution for reducing data annotation effort in 3D computer vision tasks.