MS-CAFNet: enhancing 3D point accuracy of laser scanners through a multi-stage self-attention network for applications in construction

Abstract We propose the MS-CAFNet algorithm, featuring an integrated pipeline that combines per-room geometric registration with a multi-stage convolutional network and symmetry test-time augmentation. This method enables the reduction of 3D point accuracy uncertainty associated with laser scanners (LS) in rooms that are in the shell construction phase with finished interior plastering. Thereby, it facilitates more accurate spatial measurements for geometric modeling and renovation projects. Due to different equipment limitations and environmental factors, high-end and low-end LS have positional errors. Our approach pairs high-accuracy scanners (HAS) as references with corresponding low-accuracy scanners (LAS) of measurements in identical environments to quantify specific error patterns. By establishing a statistical relationship between measurement discrepancies and their spatial distribution, we develop a correction framework that combines traditional geometric processing with targeted neural network refinement. Five-fold room-level cross-validation over 270 rooms shows a peak signal-to-noise ratio (PSNR) improvement of (7.62 ± 0.36 dB) and a mean-square-error reduction of about 68 %. The high-accuracy scanner is used as a high-accuracy working reference : the method harmonises the low-accuracy scanner to this reference by removing its systematic device-specific bias, rather than establishing a metrologically traceable uncertainty against calibrated ground-truth geometry. This approach enables low-end devices to approach the accuracy of high-end devices without hardware modifications.