Implementation U-Net for Semantic Segmentation and Perspective Transformation in Floor Tile Virtual Try-On Systems

Abstract

Conventional floor tile VTO systems rely on printed catalogs, physical samples, or marker-based AR, making it difficult for consumers to visualize ceramic products in real rooms. This study proposes a markerless VTO framework integrating U-Net floor segmentation and homography-based perspective transformation. The model was pre-trained on 5,285 SUN RGB-D images and fine-tuned on 1,338 customer-room images from PT Concord Industry, divided into training, validation, and independent test sets using a 70:20:10 ratio. Segmentation was evaluated using IoU, Dice Coefficient, Pixel Accuracy, Precision, and Recall. Perspective alignment was evaluated using four-corner Mean Angular Error (MAE) on two room images representing conditions with and without excessive furniture, while system functions were validated through black-box testing. The fine-tuned U-Net with a ConvNeXt-V2-Huge encoder achieved an IoU of 0.9483, a Dice coefficient of 0.9671, a pixel accuracy of 0.9891, a precision of 0.9675, and a recall of 0.9702. The perspective module achieved an average MAE of 5.92° and produced visually coherent tile alignment. These results demonstrate the framework's potential to support realistic floor tile visualization and reduce uncertainty during product selection.