Context. Detailed shape modeling is a fundamental task in the context of small body exploration aimed at supporting scientific research and mission operations. The neural implicit method (NIM) is a novel deep learning technique that models the shapes of small bodies from multi-view optical images. While it is able to generate models from a small set of images, it encounters challenges in accurately reconstructing small-scale or irregularly shaped boulders on Ryugu, which hinders the investigation of detailed surface morphology.
Aims. Our goal is to accurately reconstruct a high-resolution shape model with refined terrain details of Ryugu based on a limited number of images.
Methods. We propose an improved NIM that leverages multi-scale deformable grids to flexibly represent the complex geometric structures of various boulders. To enhance the surface accuracy, three-dimensional (3D) points derived from the Structure-from-Motion plus Multi-View Stereo (SfM-MVS) method were incorporated to provide explicit supervision during network training. We selected 131 Optical Navigation Camera Telescope images from two different mission phases at different spatial resolutions to reconstruct two Ryugu shape models for performance evaluation.
Results. The proposed method effectively addresses the challenges encountered by NIM and demonstrates an accurate reconstruction of high-resolution shape models of Ryugu. The volume and surface area of our NIM models are closely aligned with those of the prior shape model derived from the SfM-MVS method. However, despite utilizing fewer images, the proposed method achieves a higher resolution and refinement performance in polar regions and for irregularly shaped boulders, compared to the SfM-MVS model. The effectiveness of the method applied to Ryugu suggests that it holds significant potential for applications to other small bodies.
