Comparative Evaluation of Classical and Deep Learning-Based 3D Reconstruction Methods via Web-Based Human Perception Assessment

Abstract

The rapid advancement of 3D reconstruction technology has created significant opportunities across various fields, yet its complexity remains a barrier to widespread adoption, particularly for novice users. This research evaluates and compares traditional and modern 3D reconstruction methods to assess their effectiveness in terms of reconstruction quality and processing efficiency. Classical approaches such as COLMAP and VisualSfM are analyzed alongside the neural network-based Instant-NGP to highlight their respective strengths and limitations. A comparative study is conducted based on both objective metrics and subjective human perception, ensuring a comprehensive evaluation of their performance. Additionally, user feedback is collected to assess ease of use and accessibility, providing insights into potential improvements for broader adoption. The findings indicate that modern deep learning-based approaches offer significant advantages in speed and flexibility, while classical methods retain strengths in accuracy and consistency. To facilitate access to 3D reconstruction frameworks and ensure a more reliable user evaluation, we also incorporate a web-based interface. This eliminates the need for users to manually collect data and execute reconstruction steps, allowing them to focus solely on evaluating the final 3D reconstruction results, thereby enhancing the validity of the user survey results.