On-Device Depth Refinement for Welding Distortion Measurement

Abstract

Foundation models excel in general computer vision but struggle with precise 3D measurements in specialized industrial settings. Industrial steel plates, for instance, are often monochromatic and lack texture, challenging accurate depth inference for these models. Traditional manual measurement of out-of-plane deformation, as seen in ship hull welding, is slow and not digitally integrable. To solve this, our study proposes a novel, real-time algorithm using portable LiDAR devices. Our mission demands high computational speed and sub-1mm accuracy. To achieve these industrial objectives, we propose and apply techniques to enhance the performance of mobile LiDAR. De- formation is quantified by fitting a quadratic function and measuring the maximum perpendicular distance to a baseline. This approach significantly improves real-time performance on mobile platforms while maintaining high accuracy, effectively address- ing the limitations of general-purpose foundation models in demanding industrial en- vironments.