Huan LEI

I received my Doctorate from the University of Western Australia where I was advised by Ajmal Mian and Naveed Akhtar. My PhD was supported by the RTP scholarship of Australian Government. After graduation, I worked as a Research Fellow at the Australian National University and later at AIML, the University of Adelaide.

Email  /  Google Scholar  /  Twitter  /  Github /  CV  /  Slow Science

seek wisdom & take the blame

My research focuses on 3D vision, including deep learning for 3D perception from point clouds and meshes, surface reconstruction, efficient surface representations, and 3D generation. I am also interested in self-supervised learning
and multimodal learning for 3D applications.

This tech report explored continuous surface representations beyond the discrete geometric data formats, using the neural parametters of SDF functions, and applied them for 3D shape analysis including classfication, retrieval, and registration.

OffsetOPT reconstructs explicit surfaces from 3D point clouds without requiring normals. It generalizes the trained model to unseen data via unsupervised per-point offset optimization, achieving detail-preserving surfaces with strong scalability.

Picasso and PicassoNet++ for deep learning over heterogeneous 3D meshes. We formulate the continuous filters for mesh convolution using spherical harmonics as orthonormal basis.

By exploiting the duality between a triangle and its circumcenter, we propose to triangulates 3D point clouds into manifold meshes using a graph neural network, named CircNet. The network detects circumcenters and recovers vertex triplets of each triangle face, reconstructing a primitive mesh. Standard post-processing is then applied to convert the primitive mesh into a manifold surface.

The preliminary Picasso for geometric deep learning over 3D meshes. Please refer to our TNNLS2023 work for the latest Pytorch codes.

Fuzzy modelling makes convolutions robust to varying densities of point clouds.

Professional service

Reviewer for CVPR, TPAMI, TNNLS, TIP.