Transfer Learning on Segment Anything Model for Footwear Outsole Segmentation to Predict Footwear Slip Resistance.

Abstract

Efficient and accurate footwear outsole segmentation is essential in safety-related applications, particularly for understanding outsole-ground interaction to prevent injuries. Identifying contact regions supports applications such as slip resistance evaluation, ergonomic design, and safety standard development. However, analyzing large datasets is time-consuming and costly due to the need for precise manual annotation. The Segment Anything Model (SAM) offers versatile segmentation capabilities, but the complex patterns and textures of outsoles limit its performance, requiring task-specific fine-tuning. This study proposes a fine-tuned SAM model to improve outsole segmentation. A dataset of 40 footwear outsoles was manually annotated using a graphical tool. The developed segmentation model was used to estimate outsole-ground contact areas, which were then input into a machine learning model to predict slip resistance categories. Ground truth labels were derived from human-centered data, classifying footwear as low or high slip resistance. The fine-tuned SAM model outperformed the original, reducing loss by 8.11%, increasing IoU to 70.45%, improving accuracy from 56.90 to 78.10%, and raising the F1 score from 53.70 to 66.30%. Image quality factors such as resolution, contrast, and intensity significantly affected performance. The classification model achieved 70% accuracy and a 64% F1 score using an 80/20 train-test split. These findings support the use of fine-tuned segmentation models for improving slip resistance assessment, with important implications for preventing falls among older adults and reducing workplace injuries.