AI-Driven Virtual Dressing Room Powered by IoT Sensing and Deep Learning

Abstract

The rapid growth of online fashion platforms has created a pressing need for accurate and realistic virtual try-on technologies that enhance user confidence during garment selection. In this work, we propose an advanced AI-driven Virtual Dressing Room that leverages BlazePose-based pose estimation, U-Net and Mask R-CNN segmentation, TPS/GMM-based cloth warping, and GAN-powered try-on models such as VITON and TryOnGAN. By integrating these cutting-edge deep learning techniques, the system enables precise body landmark detection, efficient garment extraction, and highly realistic cloth fitting, thereby improving the accuracy and usability of digital try-on experiences. The incorporation of IoT components such as RFID and QR-based garment identification further enhances the system’s capability by allowing seamless interaction within physical retail environments. Through comprehensive testing on diverse user images and garment datasets, our findings demonstrate that the proposed framework significantly outperforms traditional overlay-based methods in terms of realism, alignment accuracy, and user satisfaction. The results highlight the potential of combining deep learning and IoT technologies to revolutionize virtual fashion try-on systems, offering improved decision-making for customers and reducing return rates for retailers.

Introduction

The text discusses the development of AI-driven virtual try-on technology, which allows users to digitally visualize clothing on their bodies without physical trials. Leveraging deep learning, neural networks, and machine learning, modern virtual try-on systems replicate human visual understanding by analyzing garment attributes, human pose, and body structure. The technology operates as a multi-stage pipeline—including pose estimation, segmentation, warping, and synthesis—where each stage feeds into the next, similar to layered neural networks. Advanced models like BlazePose, U-Net, Mask R-CNN, TPS/GMM, GANs (VITON/TryOnGAN), and diffusion-based architectures enable realistic cloth deformation, texture preservation, and alignment across diverse body shapes.

Key Points from Literature Review:

1. Islam et al.: Surveyed deep learning-based virtual try-on systems (VITON, CP-VTON), highlighting the role of convolutional and generative models in realistic garment transfer.

2. Rochana & Juliet: Demonstrated GANs’ effectiveness in simulating textures, contours, and seamless clothing overlays.

3. Pang et al.: Proposed FashionM3, a multimodal AI system combining visual and textual data for fashion guidance.

4. Li et al.: Developed RealVVT, a video-based try-on system that preserves garment consistency across frames.

5. Sah et al.: Explored AI fitting rooms integrated into e-commerce for personalized visualization.

6. Sanguigni et al.: Introduced Fashion-RAG for garment editing via retrieval-augmented generation.

7. Aghilar et al.: Improved pose consistency in garment transfer across body shapes.

8. Gupta & Patel: Integrated virtual try-on with recommendation engines and IoT for retail applications.

9. Karras et al.: Proposed Fashion-VDM, a video diffusion model capturing garment movements realistically.

10. Ramsey et al.: Applied generative AI for interactive and customizable fashion design.

11. Lee et al.: Highlighted geometric matching techniques (TPS/GMM) for accurate cloth warping.

12. Bi et al.: Developed wearable sensor systems for real-time pose and garment tracking.

Existing System Limitations:

• Relies on 2D overlay techniques without proper pose estimation or cloth warping, leading to unrealistic outputs.

• Lacks accurate body landmark detection, limiting garment alignment.

• Cannot perform garment deformation or warping (no TPS/GMM integration).

• Requires manual garment input, reducing automation and scalability.

• Poor adaptability to varied body types, lighting, and backgrounds.

• Computational inefficiency when attempting higher realism due to limited optimization.

Proposed System:
The proposed AI-Driven Virtual Dressing Room integrates:

• BlazePose for precise body keypoint detection (33 anatomical points).

• U-Net / Mask R-CNN for human body segmentation.

• TPS / GMM for garment warping and alignment.

• GAN-based models (VITON / TryOnGAN) for realistic synthesis.

• Optional IoT-based garment identification for automation.

Methodology & Modules:

1. User Module: Uploads full-body images or garment URLs; ensures proper lighting and posture.

2. System Module: Processes user and garment images through segmentation, pose estimation, and warping for accurate alignment.

3. Detector Module: BlazePose identifies 33 body keypoints guiding garment adaptation.

4. Classifier Module: GAN-based synthesizers generate the final realistic try-on image, blending textures and maintaining proper proportions.

Algorithm Highlight – BlazePose:
BlazePose is a deep learning framework for body keypoint estimation, detecting detailed posture coordinates (shoulders, hips, arms, torso) to guide precise garment alignment. This ensures garments warp naturally to match the user’s pose and body shape.

Conclusion

In conclusion, the development and implementation of the AI-Driven Virtual Dressing Room represent a significant a technological advancement in digital fashion and e-commerce solutions. By integrating a Blaze Pose for pose estimation, U-Net/Mask R-CNN for segmentation, TPS/GMM for cloth warping, and GAN-based synthesis models, the system demonstrates strong capability in producing realistic, accurate virtual try-on outputs. The experimental results highlight the effectiveness of combining geometric transformation models with generative networks for garment fitting, offering users an immersive and reliable try-on experience. Utilizing the computational power of cloud platforms like Google Colab and GPU-accelerated environments, the system achieves efficient processing and scalable performance. This research emphasizes the transformative potential of artificial intelligence in fashion retail, reducing product return rates, improving customer satisfaction, and bridging the gap between online and physical shopping experiences. As the system continues to evolve, integrating additional garment types, supporting 3D body models, and refining real-time try-on capabilities will further enhance its impact. Future work may lead to widespread adoption in e-commerce platforms and smart retail stores, ultimately revolutionizing how consumers interact with fashion products.

References

[1] Wang, B., Zheng, H., Liang, X., & Lin, L. (2018). Toward characteristic-preserving image-based virtual try-on network. ECCV Proceedings, 1–17. A foundational study introducing CP-VTON, enabling realistic garment transfer using geometric matching and deep learning synthesis. [2] Han, X., Wu, Z., Wu, Z., Yu, R., & Davis, L. S. (2018). VITON: An image-based virtual try-on network. CVPR, 7543–7552. This work proposed one of the first large-scale virtual try-on systems using U-Net segmentation and conditional generation techniques. [3] Minar, M. R., & Tareq, M. S. (2020). Deep learning-based virtual fitting systems: A comprehensive survey. arXiv preprintarXiv:2007.04147. Analyzes multiple deep learning-based try-on frameworks and segmentation methods. [4] Xu, X., Guan, S., et al. (2022). TryOnGAN: Body-aware GAN for realistic virtual dressing. IEEE Transactions on Image Processing, 31, 2880–2894. Introduces GAN-based person–clothing synthesis with improved texture fidelity and body alignment. [5] Jiang, N., Liu, S., & Xu, W. (2021). ClothFlow: A flow-based model for 3D-aware virtual try-on. CVPR, 10430–10439. Proposes cloth deformation flow networks enabling better garment draping and pose adaptability. [6] Redmon, J., & Farhadi, A. (2016). YOLO: Real-time object detection. CVPR Proceedings, 779–788. Essential for real-time contour extraction and garment boundary identification in try-on systems. [7] Ren, S., He, K., Girshick, R., & Sun, J. (2015). Faster R-CNN: Towards real-time object detection. NeurIPS, 91–99. Key reference for segmentation and clothing region detection tasks. [8] He, K., Gkioxari, G., Dollár, P., & Girshick, R. (2017). Mask R-CNN. IEEE TPAMI, 42(2), 386–397. Widely used for human body segmentation in virtual try-on applications. [9] Cao, Z., Hidalgo, G., Simon, T., Wei, S. E., & Sheikh, Y. (2019). OpenPose: Realtime multi-person 2D pose estimation. CVPR, 7291–7299. Core reference for human pose estimation prior to garment alignment. [10] Lugaresi, C. et al. (2019). MediaPipe BlazePose: On-device real-time body landmark detection. Google Research, 1–12.

Copyright

Copyright © 2025 Ms. M. S. Rekha, Prajwal K. S., N. Sowjanya, Prathiksha , Shivam Kumar. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.