Generalization Ability of Neural Descriptor Fields for One-Shot 3D Mapping
Sebastian Kamil Ukleja, Michael H. Hofmann, Andreas Pichler · Procedia Computer Science · 2025
Neural Descriptor Fields (NDF), have shown promising results for mapping surface details between 3D objects like grasping points in robotics. We explore the capability of NDF in generalizing point cloud data through latent encoding. The research aims to determine whether NDF can serve as a part of a foundation model for one-shot mapping processes in 3D vision, to find out if NDF can be a future key technology in the field of 3D data representation and processing and if it is able to generalize even more broadly than already shown in previous research. To achieve this, the study evaluates the encoding proficiency of the NDF by analyzing how well it captures the surface details of point cloud objects. This involves examination of each descriptor within the network, focusing on how they encode data and their relationships to one another. We find that a significant portion of the information 70.24% is shared among clustered descriptor fields, indicating a level of redundancy. However, 29.76% of the information is unique to specific fields, suggesting that the model can capture distinctive features that may be crucial for tasks requiring detailed spatial understanding and mapping non-isometric 3D shapes. Our results indicate that while some features are strongly interdependent with the 3D data, others show low correlation, suggesting that the NDF can encode latent non-spatial information.