A Study on 3D Reconstruction Techniques in Welding Scenarios

Abstract

This thesis addresses the problem of 3D perception in additive welding by proposing a unified pipeline that integrates numerical simulation, synthetic data generation, model fine-tuning, and 3D reconstruction. First, numerical simulations are used to obtain the 3D geometry of the weld and weld pool. These geometries are then imported into Blender to construct welding scenes with consistent physical scale and to generate a synthetic depth dataset with metric ground truth. On this basis, the monocular depth estimation model DepthAnythingV2 is fine-tuned for welding conditions, enabling it to recover metric depth directly from single welding images and subsequently reconstruct the 3D shape of the weld and weld pool. Experiments demonstrate that the fine-tuned model achieves a depth prediction RMSE of no more than 3 mm and a mean relative error of no more than 0.4%. The proposed method alleviates the difficulty of obtaining absolute depth in welding environments and addresses the lack of welding-specific depth datasets, providing a low-cost and scalable solution for online monitoring and quality evaluation of welding processes.