زهرا مومني نژاد

With the rapid growth of deep learning–based 3D reconstruction techniques, particularly Neural Radiance Fields (NeRF), it has become possible to generate realistic, view-dependent reconstructions from 2D images. However, practical barriers such as the need for precise an‎d time-consuming preprocessing, the high computational cost of training, an‎d the difficulty of converting implicit representations into explicit models for game engines have limited the widespread adoption of these methods in industry. This thesis proposes an integrated an‎d practical pipeline to bridge this gap, comprising three main components: • Automating an‎d stan‎dardizing preprocessing through frame extraction, photometric refinement, an‎d camera pose estimation with fallback mechanisms. •Optimizing NeRF training for constrained environments using coverage-oriented sampling strategies, staged training, an‎d memory management. •Extracting an‎d refining explicit geometry in a multi-stage process, generating optimized meshes for game engines, including curvature-guided polygon reduction an‎d level-of-detail generation. The proposed method was eva‎luated on benchmark synthetic scenes an‎d real-world data captured with consumer cameras, an‎d compared against baseline approaches. Results show that automated preprocessing improves camera pose registration success from 88–90% (manual) to over 97%. Peak GPU memory usage decreased by ~30%, while training convergence improved by 20–25%. Geometric accuracy (Chamfer distance) improved by ~35% on average, an‎d perceptual similarity (LPIPS) by ~33%. The final polygon count was also reduced by ~40%, substantially enhancing readiness for real-time use in Unity an‎d Unreal Engine. These findings demonstrate that the proposed method achieves a practical balance between reconstruction quality, computational cost, an‎d compatibility with the game content production pipeline.

بازسازي سه‌بعدي , ميدان تابشي عصبي , استخراج هندسه صريح , خودكارسازي پيش‌پردازش , بهينه‌سازي منابع محاسباتي , آماده‌سازي براي موتورهاي بازي‌سازي , كاهش چندضلعي

3D reconstruction , neural radiance fields , explicit geometry extraction , automated preprocessing , computational resource optimization , game engine readiness , polygon reduction

zahramomeninezhad

Dr. Behrouz Minaei-Bidgoli