Open Access

ARTICLE

Deep Retraining Approach for Category-Specific 3D Reconstruction Models from a Single 2D Image

Nour El Houda Kaiber¹, Tahar Mekhaznia¹, Akram Bennour^1,*, Mohammed Al-Sarem^2,3,*, Zakaria Lakhdara⁴, Fahad Ghaban², Mohammad Nassef^5,6

1 LAMIS Laboratory, University of Larbi Tebessi, Tebessa, 12002, Algeria
2 College of Computer Science and Engineering, Taibah University, Medina, 41477, Saudi Arabia
3 Energy, Industry, and Advanced Technologies Research Center, Taibah University, Madinah, 41477, Saudi Arabia
4 Lire Laboratory, University of Constantine 2 Abdelhamid Mehri, Constantine, 25000, Algeria
5 Department of Computer Science, Faculty of Computers and Artificial Intelligence, Cairo University, Giza, 12613, Egypt
6 Department of Computer Science and Artificial Intelligence, College of Computer Science and Engineering, University of Jeddah, Jeddah, 23890, Saudi Arabia

* Corresponding Authors: Akram Bennour. Email: ; Mohammed Al-Sarem. Email:

Computers, Materials & Continua 2026, 86(3), 41 https://doi.org/10.32604/cmc.2025.070337

Received 14 July 2025; Accepted 15 October 2025; Issue published 12 January 2026

Abstract

The generation of high-quality 3D models from single 2D images remains challenging in terms of accuracy and completeness. Deep learning has emerged as a promising solution, offering new avenues for improvements. However, building models from scratch is computationally expensive and requires large datasets. This paper presents a transfer-learning-based approach for category-specific 3D reconstruction from a single 2D image. The core idea is to fine-tune a pre-trained model on specific object categories using new, unseen data, resulting in specialized versions of the model that are better adapted to reconstruct particular objects. The proposed approach utilizes a three-phase pipeline comprising image acquisition, 3D reconstruction, and refinement. After ensuring the quality of the input image, a ResNet50 model is used for object recognition, directing the image to the corresponding category-specific model to generate a voxel-based representation. The voxel-based 3D model is then refined by transforming it into a detailed triangular mesh representation using the Marching Cubes algorithm and Laplacian smoothing. An experimental study, using the Pix2Vox model and the Pascal3D dataset, has been conducted to evaluate and validate the effectiveness of the proposed approach. Results demonstrate that category-specific fine-tuning of Pix2Vox significantly outperforms both the original model and the general model fine-tuned for all object categories, with substantial gains in Intersection over Union (IoU) scores. Visual assessments confirm improvements in geometric detail and surface realism. These findings indicate that combining transfer learning with category-specific fine tuning and refinement strategy of our approach leads to better-quality 3D model generation.

---

Keywords

---