Learning model in 3D

In recent years ML researchers has been making progress in recreating 3d information from a single 2d image.

Previous work on 3d reconstruction, has required multiple images from many angles. This approach can create very good 3d reconstructions given enough data, but is generally time consuming due to the large number of photos required.

In the ML approach a CNN learns the geometry of a class of objects i.e chairs, cars etc. See for example: Pytorch: Learning Efficient Point Cloud Generation for Dense 3D Object Reconstruction

Learning human 3d shape requires special consideration due to the fact that networks need to learn both the possible shapes and poses of humans, and a lot of interesting work is going on in this area.

The model above created using GraphCMR and visualized using a three.js viewer.

Drag and rotate it!

Human 3d reconstruction have many application in areas such as fitness and fashion.