Cloud
Hyundai Motor Company Namyang R&D Center: Building a 99%-Accurate Image Search and Classification System for Crash Testing
Finding a Way to Use an Overabundance of DataHyundai Motor Company’s Namyang R&D Center faced growing challenges in managing and utilizing the massive volume of image data generated during vehicle crash safety tests. Each single crash test produces high-speed camera footage, tens of thousands of high-resolution images, and measurement data recorded by thousands of sensors. However, image data was stored across multiple folders without a consistent classification standard.To find deformation cases for a specific vehicle model and viewing angle, researchers often had to ask colleagues who had conducted the test or manually open countless folders and check thumbnails one by one. Preparing a single report required an average of 20–30 minutes just to locate and organize relevant images, creating a clear bottleneck in research efficiency and data reuse.“We had a lot of data, but no efficient way to process it. To find images of a specific vehicle and angle, we had to ask colleagues or manually check thumbnails one by one.”
— Researcher, Hyundai Motor Company Namyang R&D CenterTo overcome these limitations and improve research efficiency, Hyundai Motor Company’s Namyang R&D Center decided to build an AI-based image classification and search system. Elice was selected as the technology partner to develop a solution optimized for the crash test environment.
An AI Search Engine That Finds Photos Using PhotosIn crash test analysis, researchers need to quickly compare images from similar cases. When this process relies on human memory and manual searching, data utilization inevitably becomes inefficient—no matter how much data is accumulated. What researchers needed was the ability to search for images using images: an environment where images matching specific conditions could be instantly retrieved and compared.The challenge was that conventional image search methods were not suitable for the crash test context. For example, when submitting a diagonal-view image of a vehicle, results often prioritized images with similar colors or external appearances rather than the same viewing angle, making it difficult to find analytically relevant cases. To address this, Elice designed a search system that focuses not on visual similarity but on camera angle (view) and composition.“Conventional general-purpose models are trained mainly on visual features such as shape and color, so they struggle to classify images based on camera angle, which is critical in crash testing. Hyundai needed a model where a diagonal-view image of a white SUV would return diagonal-view images—even if they were of a red compact car.”
— Suin Kim, CRO (Chief Research Officer), EliceAs a result, researchers can now explore similar cases based on meaningful analytical criteria—camera angle—without being affected by vehicle type or color.
Precision Search Down to the Finest DetailsIn crash test analysis, not only the overall vehicle shape but also subtle changes—such as bumper curvature or damage to a lamp—often provide critical insights. Elice implemented a focus search feature that allows users to select a specific region of interest within an image, assigning greater weight to that area when retrieving similar images.This feature captures both the global context and fine-grained details of an image, enabling highly precise searches—like finding a needle in a haystack within a massive archive. Researchers can now selectively retrieve deformation cases of specific components at specific angles.
High Accuracy and Stability for Real-World UseElice began the project by deeply understanding how Hyundai’s researchers actually work—how crash test data is generated, stored, searched, and used in practice. Based on this analysis, Elice designed a custom labeling framework that accounted for camera angles, vehicle parts, and test conditions. The goal was not simply to build a technically feasible system, but one that researchers could genuinely use in their daily work.For Namyang R&D Center, the objective was not an experimental model, but an automation system deployable in real operations. Model accuracy was therefore a critical requirement. If accuracy were insufficient, human revalidation would still be required, limiting the benefits of automation.With data reliability as the top priority, the Elice AI team set a high accuracy target suitable for real-world use. As a result, the image classification system achieved an accuracy level of approximately 99%.“We initially thought around 95% accuracy would be sufficient, but Elice proposed targeting over 99% from the beginning.”
— Researcher, Hyundai Motor Company Namyang R&D CenterBeyond technical development, Elice also took on the role of an AI project manager (AI PM) to ensure the project’s successful execution. In large-scale AI projects, effective communication with stakeholders, schedule management, priority setting, and strategic alignment are just as important as model performance. Managing these elements systematically is essential for delivering a solution that can be applied in real-world operations.“In large-scale AI projects like this, the AI PM role—overseeing development, communication, and scheduling—is critically important. We systematically managed everything from communication with practitioners to modeling decisions, direction alignment, and priority management, enhancing both the completeness and stability of the project.”
— Suin Kim, CRO (Chief Research Officer), EliceDeveloping High-Precision AI Models Through Large-Scale TrainingOnce the labeling framework was established, Elice conducted large-scale training using its high-performance computing infrastructure. Through tens of thousands of iterative training cycles, the team continuously improved model accuracy and applied data augmentation techniques to ensure consistent performance across diverse shooting conditions and vehicle types.The model was ultimately advanced beyond simple visual similarity matching to a level where it could understand spatial context and measure semantic similarity.“Elice researched multiple approaches with us and significantly refined the system. We were very satisfied with the results.”
— Researcher, Hyundai Motor Company Namyang R&D CenterSecure and Stable System Integration and DeploymentAfter model development was completed, the system was integrated with Hyundai Motor Company’s existing research infrastructure. Because the system handles highly sensitive research data, security and stability were non-negotiable prerequisites.Elice strictly adhered to Hyundai’s security guidelines when designing the enterprise cloud environment and built an intuitive interface that researchers could easily use. Even after deployment, Elice continuously collected and reflected user feedback to further improve usability in real-world research environments.From Images to Documents: A Scalable Multi-Data Understanding Architecture
Beyond building an image-based search and classification system, Elice also considered a technical architecture scalable to broader research data environments. In real-world research settings, not only images but also Excel files, PowerPoint documents, and handwritten materials serve as critical analytical assets.Elice has developed 'Helpy Vision', a model equipped with proprietary Document Parser technology that automatically recognizes document layout elements and converts tables, formulas, and charts into structured data. Helpy Vision understands different document components—such as paragraphs, tables, figures, and formulas—in reading order and transforms them into structured data that AI systems can utilize.Built as a Vision-Language Model (VLM), Helpy Vision jointly interprets images and documents, enabling integrated data understanding that goes beyond image search to include document context. This capability is designed to expand toward Vision-Language-Action (VLA) models in the future, enabling applications in physical AI environments.Advancing AX with EliceElice supports AI system development end to end—from data organization and model development to real-world deployment—while ensuring not only accuracy but also operational stability and practical applicability. Even in highly specialized industrial data environments such as crash testing, data only becomes an asset when it can be effectively used.If you are looking to build an AI environment that allows your organization to leverage large-scale data more quickly and accurately, start with Elice.
