• Today, a workshop was held on advances in simulation and synthetic generation of annotated data for autonomous vehicles.
• Virtual scenarios, generative AI, and large-scale data analysis are setting the research trend.

A Coruña, December 20, 2024.- CITIC of UDC has been the meeting point to address advances in simulation and synthetic generation of annotated data for autonomous vehicles. The current situation, potential, possible research lines, and future collaborations on autonomous driving were the focus of the workshop organized by the Computer Graphics & Visual Computing (XLab) group in collaboration with CITIC.

Researchers from the project Automatic Data Annotation by Semi-Supervised Learning and Procedural Generation of Augmented Scenarios (ADA), part of the Knowledge Generation R&D Projects call by the State Research Agency of the Ministry of Science, Innovation and Universities, presented results, reflections, and contributions from the subproject Procedural Generation of Augmented Scenarios with Automatic Data (PGASADA). The event included researchers from CITIC of UDC, the Polytechnic University of Catalonia, and the Jaume I University of Castellón.

Projects: Overview and Advantages

One of the works presented by José A. Iglesias Guitián, called UrbanSyn, is an open synthetic dataset offering photorealistic driving scenes with over 7,500 images. Designed to bridge the gap between synthetic and real domains, it offers, according to Iglesias, “diversity and quality in multiple driving scenarios, with a high degree of realism achieved through advanced ray tracing and denoising techniques.” UrbanSyn is open for both research and commercial purposes.

The workshop also addressed the opportunities offered by connected, cooperative, and automated mobility and the transition from technology as an end to technology as a tool. Margarita Martínez Díaz from the Polytechnic University of Catalonia presented data on the national PLATOON project, which defines management strategies for platoons of autonomous vehicles on highways, and the European CulturalRoad project, which seeks to develop a methodology to evaluate the equity of CCAM-based solutions, focusing on geographical and cultural diversity.

The generation of synthetic datasets was another topic covered in the presentation HighwayScapes: Generation of per-pixel metadata in 2D images for highways, presented by José Ribelles from Jaume I University. He introduced a new system to create annotated 2D images from 3D virtual worlds, specifically designed for highways.

Synthetic data generated in 3D virtual scenarios were also discussed by Manuel Silva from CITIC as a common alternative to create pixel-level labeled image datasets, useful for training neural perception models such as semantic segmentation.

Another line developed as part of the project Automatic Data Annotation by Semi-Supervised Learning and Procedural Generation of Augmented Scenarios (ADA), in which CITIC researcher Omar A. Mures works, is the use of generative AI as an alternative for generating synthetic datasets for autonomous driving. His work explores the use of generative AI models to automatically create labeled synthetic data. These models allow for the introduction of diversity, such as weather conditions or fine details, overcoming the limitations of traditional simulations, which are often costly.