Browsing by Subject "3D computer vision"

Now showing 1 - 2 of 2

Improving 3D shape generation by shape space refinement and details recovery
(2022-05-10) Sun, Bo (M.S. in computer science); Liu, Qiang (Ph. D. in computer science); Wang, Atlas
In this work we discuss two novel perspectives to improve 3D shape generation. The first perspective is to improve the local rigidity of the shape space in shape generation from a latent vector, while the second perspective is to use patch copy to do details recovery in shape completion. In the first improvement, we introduce an unsupervised loss for training parametric deformation shape generators. The key idea is to enforce the preservation of local rigidity among the generated shapes. Our approach builds on an approximation of the as-rigid-as possible (or ARAP) deformation energy. In the second improvement, we introduce a data-driven shape completion approach that focuses on completing geometric details of missing regions of 3D shapes. Our key insight is to copy and deform the patches from the partial input to complete the missing regions. This enables us to preserve the style of local geometric features, even if it is drastically different from the training data.
Robust framework for 3D synchronization
(2023-12) Sun, Yifan; Huang, Qixing; Hua, Gang; Biswas, Joydeep; Vouga, Etienne
The 3D synchronization task stands for weaving a set of 3D objects and their representations into a self-consistent global configuration. With an increasing number of large 3D datasets becoming accessible, efficient synchronization methods are required for applications in 3D vision. In this work, we propose a simple and efficient framework for solving synchronization tasks over a set of 3D objects. We start with a general map synchronization over symmetric 3D objects and present a joint map representation for solving a globally consistent map. Next we propagate to the 3D registration problem and propose a discretized spectral method to compute consistent global poses of symmetric partial objects. Next, we handle the setting with multiple estimation over a single edge through our diffusion-and-clustering method over continuous pose distributions. After the global phase, we shift our focus to the quality of multi-view representation and pairwise maps. For the quality of multi-view representation of 3D objects, we present a learning based method which efficiently solves the view selection problem for covering indoor scenes. Finally, we introduce a data-driven method to improve the accuracy of pairwise transformations over the transformation graph for the multi-piece 3D assembly task with small overlaps between pieces, which extracts shape features from local geometry for surface classification and piece matching.