GIST Develops Groundbreaking Technology to Dramatically Reduce 'Data Annotation' Workload

Professor Kyubin Lee Combines Semi-Supervised and Active Learning in the Department of AI
Workload Reduced to 46% Compared to Conventional Methods

The Gwangju Institute of Science and Technology (GIST) announced on September 8 that Professor Kyubin Lee’s research team from the Department of AI has developed a technology that significantly reduces the workload of ‘data annotation,’ a process in which the locations and areas of objects in images needed for AI training are marked to create ground-truth data.

Group photo of Professor Kyubin Lee's research team from the AI Department at GIST. Provided by GIST.

This technology expands the range of data that AI can learn from autonomously. It automatically processes results where the AI is relatively confident in its judgment, while focusing human review only on results that are difficult for the AI to judge. As a result, training data for AI can be built in less time and with less effort compared to the conventional method where humans must manually check and correct every piece of data.

For AI to be used in real-world environments like self-driving vehicles or underwater exploration robots, large amounts of image data depicting a variety of situations are required. In particular, to train AI to accurately recognize objects within images, humans must manually indicate the ‘correct answers’ for each image—a process called data annotation.

For example, if there are several cars in a single photo of a road, each car must be individually identified and marked. The ‘instance segmentation’ method, which the research team focused on, goes a step further by identifying and marking the area of each object on a pixel-by-pixel basis within the image. This makes processing each image a labor-intensive and time-consuming task.

To alleviate this burden, ‘active learning’ methods have previously been used. In this approach, the AI selects images that are useful or difficult for learning and requests human annotation for those selections. However, this method mainly trains the AI on data already annotated by humans, resulting in a limitation where many unannotated images are insufficiently utilized.

Additionally, the confidence the AI assigns to its own judgments does not always match actual accuracy. This discrepancy can become larger when there are either too many or too few specific types of objects in the data.

To address these limitations, the research team combined semi-supervised learning and active learning, and developed a new data annotation system that calibrates the AI’s confidence according to the characteristics of the data. The researchers’ method can be understood as involving three main stages.

First, images that have not been annotated by humans are also used for AI training. Next, the AI selects images that are either more difficult to learn from or are likely to provide new information, and requests human annotation for these images. Lastly, prediction results generated by the AI are sorted by confidence level: highly confident results are automatically accepted without human review, while ambiguous results are sent to humans for verification.

The operation process of the AI training data annotation system developed by the research team. Provided by GIST

In particular, rather than using AI-generated confidence scores as-is, the research team adjusted confidence by considering object types and frequencies found in images. This is important because, even if the AI reports similar confidence scores for both common and rare objects, actual accuracy can differ significantly between these categories.

To verify the system’s effectiveness, the team conducted experiments using the real-world road environment image dataset ‘Cityscapes,’ commonly used in self-driving research. Five participants experienced in data annotation each performed annotation tasks across 150 urban road images in five rounds, comparing the following approaches: manual annotation, the conventional method (combining active learning and AI-assisted annotation), and the method developed by the research team. As a result, with the team’s method, the time spent by humans on data annotation was reduced by 46.3% compared to the conventional approach and by 68.4% compared to manual annotation.

At the same time, the AI’s image recognition performance improved. In the final stage, the team’s method recorded a mask Average Precision (mask AP) of 30.3, which was 1.8 points higher than the 28.5 achieved by the conventional approach and 4.5 points higher than the 25.8 recorded by the manual annotation approach.

Professor Kyubin Lee explained, "The core of this research is allowing the AI to handle data that it can process while humans focus only on parts requiring human judgment. We expect this to help reduce the time and burden of data construction in fields that constantly require new training data for unique environments, such as autonomous driving or underwater exploration robots."

This study was supervised by Professor Kyubin Lee of the Department of AI, with Joungwon Kim, a doctoral student, contributing as the first author. The research was supported by the Ministry of Science and ICT and the Institute of Information & Communications Technology Planning & Evaluation (IITP) under the 'Human-Centered Core AI Technology Development' program. The results were published online on July 15, 2026, in the international journal 'Intelligent Systems with Applications' in the field of information systems.

This content was produced with the assistance of AI translation services.