Scholarlink

<p>Three-dimensional object reconstruction from images is a core area in computer vision, primarily aimed at deriving the geometric structure of an object from images. A prominent technique within this domain is multi-view photometric stereo, which leverages multiple images captured from different camera viewpoints under varying illumination conditions to accurately infer surface geometry. This paper proposes a novel approach to 3D reconstruction by utilizing the displacement of spatial points to train an implicit network. The loss function is principally based on surface gradients and the Shape from Silhouettes, aiming to address the challenge of 3D reconstruction by multi-view photometric stereo using a neural inverse network. The main concept of this method is to train a multilayer neural network to determine the surface of an object. This technique integrates surface orientation data and surfaces derived from image silhouettes to form a robust loss function. It leverages the movement of points along direction vectors, utilizing surface gradients and shape from silhouette domain, focusing exclusively on the essential information required to train the neural network. A detailed experimental section will be presented in this paper to showcase the obtained results, highlighting the high performance of our new approach. We demonstrate that the proposed method outperforms existing techniques in terms of quality and processing time across several tests. Our code will be available on github after acceptance at : <a href="https://github.com/lyabada/SPSN/.">https://github.com/lyabada/SPSN/.</a></p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

SPSN-MVPS: Shifting points, silhouettes and neural inverse for solving multiview photometric stereo

  • 期刊论文

  • 发表日期：

    2025年6月5日

查看全文
  • Lyes Abada,
  • Tarek Gacem,
  • Aimen Said Mezabiat,
  • Saad allah Bourzam,
  • Omar Chouaab Malki,
  • Mohamed Mekkaoui

摘要

Three-dimensional object reconstruction from images is a core area in computer vision, primarily aimed at deriving the geometric structure of an object from images. A prominent technique within this domain is multi-view photometric stereo, which leverages multiple images captured from different camera viewpoints under varying illumination conditions to accurately infer surface geometry. This paper proposes a novel approach to 3D reconstruction by utilizing the displacement of spatial points to train an implicit network. The loss function is principally based on surface gradients and the Shape from Silhouettes, aiming to address the challenge of 3D reconstruction by multi-view photometric stereo using a neural inverse network. The main concept of this method is to train a multilayer neural network to determine the surface of an object. This technique integrates surface orientation data and surfaces derived from image silhouettes to form a robust loss function. It leverages the movement of points along direction vectors, utilizing surface gradients and shape from silhouette domain, focusing exclusively on the essential information required to train the neural network. A detailed experimental section will be presented in this paper to showcase the obtained results, highlighting the high performance of our new approach. We demonstrate that the proposed method outperforms existing techniques in terms of quality and processing time across several tests. Our code will be available on github after acceptance at : https://github.com/lyabada/SPSN/.