Object Centric Point Sets Feature Learning with Matrix Decomposition

Zijia Wang, Wenbin Yang, Zhisong Liu, Qiang Chen, Jiacheng Ni, Zhen Jia; Proceedings of the Asian Conference on Computer Vision (ACCV) Workshops, 2022, pp. 251-264

A representation matching the invariance/equivariance characteristics must be learnt to rebuild a morphable 3D model from a single picture input. However, present approaches for dealing with 3D point clouds depend heavily on a huge quantity of labeled data, while unsupervised methods need a large number of parameters. This is not productive. In the field of 3D morphable model building, the encoding of input photos has received minimal consideration. In this paper, we design a unique framework that strictly adheres to the permutation invariance of input points. Matrix Decomposition-based Invariant (MDI) learning is a system that offers a unified architecture for unsupervised invariant point set feature learning. The key concept behind our technique is to derive invariance and equivariance qualities for a point set via a simple but effective matrix decomposition. MDI is incredibly efficient and effective while being basic. Empirically, its performance is comparable to or even surpasses the state of the art. In addition, we present a framework for manipulating avatars based on CLIP and TBGAN, and the results indicate that our learnt features may help the model achieve better manipulation outcomes.