3D character reconstruction from 2D orthogonal images

Kok, Yung Thow (2025) 3D character reconstruction from 2D orthogonal images. Final Year Project, UTAR.

PDF Download (3870Kb)

Abstract

The demand for 3D content in games and films has driven the need for more efficient character modelling workflows. Traditionally, 3D character creation from concept art requires time-consuming manual modelling and sculpting, followed by retopology, texturing, and rigging. These processes may need weeks to complete for a single character. While recent AI-based 3D generation models offer faster alternatives, but lack of fine control over the output and are often requires extensive post processing to ensure the quality of the mesh. This research is motivated by the need to support artists in traditional 3D character modelling workflow that refine the details of mesh from primitive shapes. Instead of fully automating the modelling process, this project aims to introduce an algorithm that reconstruct primitive cubes of stylized 3D character from front, left, back and right orthogonal views images by integrating interactive image segmentation and blender scripting method. This enables a rapid starting point for geometry detail adjustment using box modelling or digital sculpting techniques. The project objectives are to: (1) review the existing state-of-arts for solving 3D generation and reconstruction, (2) design an algorithm capable of reconstructing primitive shapes of 3D characters from four orthogonal images, optimized for limited computational resources. This research contributes a flexible and artist-friendly reconstruction algorithm that capable of generating part-level primitive shapes, which is efficient for generating high-fidelity mesh using box modelling and digital sculpting techniques. The algorithm can generate primitive shape from at least two orthogonal view images. The proposed algorithm is organized into modular stages, allowing artists and developers to further improving the pipeline by integrating computer vision, AI models and 3D modelling and scripting techniques. Finally, the model’s performance will be evaluated based on topology quality, editability, and usability. This work not only improves the character modelling pipeline, but also supports future studies in 3D reconstruction and digital content creation,

Actions (login required)