Accuracy Assessment of Dense Point Cloud Generated by Deep Learning and Semiglobal Matching

This study assesses two techniques for generating point clouds based on dense image matching (DIM): semiglobal matching (SGM) implemented in Trimble INPHO MATCH-3DX software, and a deep-learning algorithm Pyramid Stereo Matching Network (PSMNet). The PSMNet was trained using three datasets with both automated driving and aerial scenes. Two other distinctive sites were selected to assess the accuracy against LiDAR data. The study found inaccuracies in the PSMNet point clouds and suggested that SGM could potentially result in a better outcome. However, for flat slab or ground surface, its root-mean-square error is better than SGM. The analysis showed that the SGM analyses favorably remove points on vertical surfaces due to occlusion, while PSMNet incorrectly extrapolate them with slopes. Furthermore, the assessment identified the potential to improve PSMNet by using more or in-distribution training dataset for test (unseen) areas.