cv

Ph.D

University of Southern California

Undergraduate

Tsinghua University

• GPA: 3.83
• Scholarship
  • Excellent International Student, sponsored by the Beijing Government

NeRFPrior: Learning Neural Radiance Field as a Prior for Indoor Scene Reconstruction

• We propose a 2 stage training pipeline for faster and more detailed 3D scene reconstruction. On the first stage, we use a voxelized neural network to approximate the radiance field. On the second stage, we input the coarse radiance field as prior into a SDF-based Nerf and get a consistent and finegrained field for mesh reconstruction.
• Experiments show that our method achieve better reconstruction performance with less training time, compared with previous method such as Neus and unisurf. Meshes retrieved from our method have flatter wall and floor area and more accurate shape for small crafts such as bowls and spoons on table.
• This work is submitted!

Domain Adaptation on Point cloud Completion

• We propose to use structure as a guide for point cloud completion. Given a partial scan, we first predict its coarse cuboid structure using conditional GAN. Next, we refine the coarse cuboid prediction and output complete point clouds.

Learning structure deformation using cuboid abstraction

• We propose a method to learn possible variation for certain kind of human-made objects unsupervisedly. We first extract the cuboid structure for each object. Then we learn several meta variations for the predicted structure by deforming the object to other objects in the set.

Chrome Recorder

• A recorder to caputer user action on websites. Used for data collection by Tsinghua NLP team.

MonoRCore

• This is my graduation design program. A modular Rust operation system with scheduling and page replacement.