Embedded Files
Aim: Train a neural network to use a single 2D image to generate a full 3D point cloud
Over 3000 images of 3D models rendered in Blender
Over 3000 images of 3D models rendered in Blender
A UNET architecture was trained
A UNET architecture was trained
with the chamfer distance as a loss function
Point Clouds were predicted by the network
Point Clouds were predicted by the network
3000 images: The models used to generate both ground truth point clouds and rendered images were from the authors of Semantic Parametric Reshaping of Human Body Models
UNET: A UNET is a type of Neural Network shaped like a U (as shown above), the SkipNet connections can help minimise problems such as gradient vanishing
Point Clouds: A point cloud is a collection of 3D coordinates represented by a point
Chamfer Distance: The chamfer distance is a metric used to determine the similarity between to Point Clouds. I have made an explanatory video here
Google Sites
Report abuse