Vera van der Lelij - Adjusting 2D prediction models to be usable for 3D objects

Discover predictive 3D printing techniques using 2D image predictions and overcome point cloud challenges with symmetric functions.

Key takeaways

Vera van der Lelij introduces the concept of using 3D objects and 2D images to make predictions.
Data comes from various sources, including 3D printing.
Many neural networks can work with images, but fewer can work with 3D objects.
The speaker works on a project at PIPL to predict object shape changes in 3D printing.
Permutation and variance in point cloud data can lead to problems with mapping transformations.
Using symmetric functions, the speaker and her colleagues align and transform objects, then create a global feature vector to predict object shape changes.
Open3D and other libraries help with point cloud manipulation and feature extraction.
The speaker introduces point cloud, voxel grid, and mesh object data types, each with different properties.
They explore the connection between point cloud representation and forecasting, but mention the importance of aligning objects.
Classifying or forecasting can also be used, but specific techniques apply to each task.
The speaker also notes that there’s a connection to 3D printing and STL file format.

More talks