Portal de Periódicos da CAPES

3D Avatar Reconstruction Using Multi-level Pixel-Aligned Implicit Function

2024; Springer International Publishing; Linguagem: Inglês

10.1007/978-981-99-9442-7_20

2367-3370

Shreedhar I. Muttagi, Vaishnavi Patil, Pooja P. Babar, Ritvik Chunamari, Uday Kulkarni, Satish Chikkamath, S M Meena,

Advanced Vision and Imaging

In recent years, there has been significant interest in 3D avatar reconstruction from images, driven by its applications in gaming, entertainment, and augmented reality. This paper introduces a novel approach that utilizes the Fast Super-Resolution Convolutional Neural Network (FSRCNN) and multi-level pixel-aligned implicit function (ML-PIFu) framework to reconstruct high-fidelity 3D avatars from a single input image. Our method integrates the FSRCNN with a two-module ML-PIFu pipeline, combining global geometric information and local fine details. The FSRCNN enhances the input image quality before it is processed by the ML-PIFu framework. The coarse-level module of ML-PIFu improves the image quality, capturing the overall structure and shape of the human form, while the fine-level module incorporates intricate details for enhanced output. Through the use of neural networks and an end-to-end training strategy, our approach achieves accurate and precise reconstructions without requiring multiple images or complex camera setups. The potential applications of our approach are extensive, including the creation of personalized avatars for telepresence, Virtual Reality (VR), Augmented Reality (AR), anthropometry studies, and virtual try-on experiences. By leveraging our approach, these applications can benefit from enhanced accuracy, realism, and versatility, thereby opening up new possibilities in the realm of digital human representation and interaction.