This work is supervised by Prof.Li Cheng, cooperated with PhD.Ji Yang, University of Alberta.

Previous works

Before I join in this project, the team have done these works. The paper was published on ECCV 2022.

  • Render a skeleton prediction approach by utilizing the deep implicit functions.
  • Build an automated pipeline to tackle the entire process of 3D reconstruction, rigging, and animation, all from single-view RGB images.
This two-stages architecture for 3D Rigging. The first stage on the left adopt the deep implicit method to reconstruct the 3D mesh from a single image. The right part was used to generate an 3D probability field, which aims to predict object's skeleton.
This image shows the whole process, including 3D reconstruction, skeleton, rigging, and animated result.

Works now

We are trying to improve the previous work. Inspired by NeRF, we are making a great effort to adapt our two-stages network for an end-to-end pipeline. In this work, I design a neural network to extract the canonical feature from different views of an identical 3D model, which is pre-trained on rendered ShapeNet. We have got a little more accurate result than SOTA. The paper and code are on the way soon.