Construction Domain-Adapted Synthetic Data Generation

A construction computer-vision object recognition model often fails repeatedly in real-world deployments. An adaptive data generation framework for training can use these repeated failures as signals to organize the causes of performance degradation and to refine the data composition, thereby stabilizing the model over time. In practice, there are many differences between training data and the operational environment, such as illumination, viewpoint, occlusion, background complexity, and changes in target characteristics. These discrepancies tend to concentrate false positives and false negatives under specific conditions and widen the domain gap. As a result, simply collecting more data may fail to capture the problematic conditions sufficiently or may lead to inefficient improvements, creating a vicious cycle in which performance degradation continues to recur.

To mitigate this vicious cycle, this study does not merely accumulate failure cases. Instead, it summarizes the conditions under which failures occur in a form that can be systematically analyzed, and it feeds this information back into data refinement and training in a closed-loop manner. Specifically, based on error patterns observed during operation(A), the framework categorizes vulnerable situations and qualitatively and quantitatively identifies the conditions that drive performance degradation(B,C). It then adjusts the dataset composition so that these vulnerable conditions are adequately represented, and, when necessary, applies complementary strategies including synthetic data to fill gaps in the data distribution(D). Finally, by continuously updating the dataset and training settings using feedback from repeated performance evaluations, this study aims to establish a data-centric pipeline that maintains stable performance under the highly variable conditions of construction sites.