Real-Time AI-Driven Interior Design System with Augmented Reality

Interior design has traditionally relied on two-dimensional visualizations and physical prototypes, which limit interactivity and make it difficult to preview design changes in real time. This paper presents a real-time, AI-driven interior design system that combines semantic segmentation with markerless augmented reality (AR) to let users place virtual furnishings inside a live camera view of their own room. A U-Net segmentation network with a MobileNetV2 encoder was trained on the ADE20K indoor-scene dataset to identify walls, floors, and furniture, and was converted to the Open Neural Network Exchange (ONNX) format so that it could run natively inside the Unity engine. The segmentation output is combined with depth-gradient and surface-normal estimation to support plane detection, on top of which the Unity AR Foundation pipeline (AR Session, AR Plane Manager, AR Raycast Manager, and AR Anchor Manager) anchors and renders virtual rugs and furniture onto the detected surfaces. The proposed U-Net model, trained on 5,000 augmented indoor images, is compared against a baseline U-Net trained on the PASCAL VOC 2007 segmentation dataset. The ADE20K-trained model reached a training dice coefficient of 0.80 and a validation dice coefficient of 0.58, while the VOC-trained baseline achieved a validation accuracy of 67.65% and a validation Intersection-over-Union (IoU) of 0.184. Field tests in living-room and library environments show that the segmentation-driven plane detection pipeline can anchor virtual rugs to real floor surfaces under everyday lighting conditions, and qualitative evaluation indicates clear gains in user engagement, decision-making support, and time-saving compared with static, non-interactive design tools. The paper also reports the outstanding gap between training and validation performance, indicative of overfitting, and outlines optimization and user-study directions for future work.