3D reconstruction of toys based on adaptive scaled neural radiation field

With the rapid development of computer vision technology, 3D reconstruction of toys under single-view conditions still faces significant challenges in terms of detail loss and color distortion. For this reason, this article proposes an adaptive scale neural radiance fields (AS-NeRF) model to enhance...

Full description

Saved in:
Bibliographic Details
Main Authors: Jiajun Zou, Shaojiang Liu, Feng Wang, Weichuan Ni, Shitong Ye
Format: Article
Language:English
Published: PeerJ Inc. 2025-07-01
Series:PeerJ Computer Science
Subjects:
Online Access:https://peerj.com/articles/cs-3053.pdf
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:With the rapid development of computer vision technology, 3D reconstruction of toys under single-view conditions still faces significant challenges in terms of detail loss and color distortion. For this reason, this article proposes an adaptive scale neural radiance fields (AS-NeRF) model to enhance the accuracy and realism of 3D toy reconstruction. The method constructs a multi-task feature extraction network based on the Vision Transformer, which simultaneously extracts and fuses multidimensional features such as texture, shape, color, and depth through a task dynamic modulation mechanism and a dynamic adapter layer, providing a rich and accurate contextual feature representation. The NeRF model is enhanced to incorporate an adaptive scaling mechanism that dynamically optimizes rendering sampling accuracy according to the local complexity of the scene. Spectral sensing techniques are integrated to reproduce the true colors of materials accurately. Finally, the conditional diffusion model is deeply integrated with NeRF, and high-dimensional conditional vectors are used to guide the inverse diffusion process in generating unobserved images with consistent geometric structure and physical properties. Experiments on the Co3D toy dataset demonstrate that AS-NeRF significantly outperforms existing mainstream methods in terms of peak signal-to-noise ratio (PSNR), structural similarity (SSIM), loss of perceptions (LPIPS), and Chamfer distance, thereby verifying the validity and advantages of the proposed method for high-quality toy 3D reconstruction tasks.
ISSN:2376-5992