Chair reconstructed by Deep Neural Network

Being provided:
1. a point cloud normalized inside the unit cube
2. a set of query points (that is, uniformely sampled points inside the unit cube, aka ambiant points)
3. the value of the signed distance function for those query points
a deep neural network (Convolutional Occupancy Network) learn toe modelize the signed distance function.
Once trained, it can take new point clouds and output their distance map (occupancy probabilites within a grid of fixed resolution), which can then be used to reconstruct meshes, using the MISE algorithm (Multi-scale Iso-surface Extraction). This is the result of such a process.