Speaker：Shiqing Xin（Shandong University）

Time：2022-11-17, 10:30

Location：Tencent Meeting ID：581-4735-3770（No Password）

Abstract:

3D reconstruction is the core technology of Metaverse content construction and the key link of reverse engineering. The report explores how to reconstruct high-fidelity 3D shapes from low-quality point clouds. The core idea is to solve the contradiction between "low data quality" and "high target requirements" based on weak prior knowledge. By absorbing manifold priors in explicit reconstruction, the ambiguity problems caused by thin plates and thin tubes can be resolved; by absorbing the priors with obvious feature lines of CAD models, CAD models with regular structure and clear feature lines can be reconstructed; by absorbing MLP The global prior of the network basis function, self-supervised learning is performed, and the geometric model with high fidelity is restored when the point cloud is sparse, noisy, and severely missing. The report will also summarize the crux of the problems faced by 3D reconstruction and look forward to future development trends.