CVPR 2023 – OReX: Object Reconstruction from Planar Cross-sections Using Neural Fields

In this episode we discuss OReX: Object Reconstruction from Planar Cross-sections Using Neural Fields
by Haim Sawdayee, Amir Vaxman, Amit H. Bermano. The paper presents OReX, a method for reconstructing 3D shapes from planar cross-sections using a Neural Field as the interpolation prior. The trained neural network estimates the inside/outside function of a given 3D coordinate and induces smoothness and self-similarities. The method addresses the challenge of high-frequency details by using an iterative estimation architecture and a hierarchical input sampling scheme, allowing the training process to focus on high frequencies at later stages. The paper demonstrates the method’s robustness, accuracy, and scalability and reports state-of-the-art results compared to previous approaches and recent potential solutions.