Speeding up the annotation process in semantic segmentation industrial applications

Researchers have demonstrated that unsupervised computer vision algorithms can significantly accelerate the labeling process for semantic segmentation tasks in industrial materials science, reducing labeling time by approximately 78%.

A recent study showcased the potential of unsupervised algorithms in improving data annotation efficiency for complex semantic segmentation problems. The annotation process for semantic segmentation tasks is time-consuming and prone to human errors, particularly when dealing with high-resolution images such as those used in microstructure characterization in materials science^[1]. The study used a dataset comprising large images of dimensions 1280x959 and 960x703, further increasing the complexity of the annotation task. By leveraging unsupervised computer vision algorithms, the researchers were able to reduce the labeling time from 170 hours to 37 hours. The study also created and shared the largest public steel microstructure segmentation dataset to date, available under MIT License with permanent DOI. Additionally, the authors of a related study achieved a composite mean Intersection-over-Union (mIoU) of 69.73% in the ICRA 2026 GOOSE 2D Fine-Grained Semantic Segmentation Challenge, using the Segment Anything Model 3 (SAM3) as a visual foundation model and contributing a self-distillation scheme that re-uses SAM3 as a teacher on classes where it outperforms their own model^[2].