Abstract 3D deep-learning based models have been proposed to assist in the digitalization process of as-is infrastructure. Despite their outstanding performance to process large amounts of data, research on interpreting their execution rather than improving their outcome has been scarce. To further characterize the execution of these models, in this study we seek to identify influential parts of the input that contribute the most to a given model decision in the context of 3D semantic segmentation. Considering the intrinsic nature of 3D data and the existing basic characterization techniques, we explore variations of the rules to backpropagate the model prediction signal back to the input. On the qualitative side, we have found that the techniques implemented point to learned patterns that vary depending on the model architecture. We show that the generated explanations serve also to identify spurious learned patterns that would otherwise be unnoticed if only analyzing the model predictions. Our results serve then as a baseline comparison for further research into 3D explanation techniques.
Cervantes et al. (Fri,) studied this question.