Deep Learning meets 3D Technology

A new level in industrial image processing

Industrial image processing has reached a new level in understanding image, audio, and text data. 3D point clouds can be processed using deep learning.

The two most common representations of 3D data are point clouds an meshes. However, neither one is based on a regular grid. As the performance of deep neural networks usually benefits from regular structures, applying them to raw data is difficult.

To analyse this kind of data in neural networks, two approaches are currently established:

  1. Transferring existing scene information into a regular representation such as 3D volume grids or multi-view-images in a pre-processing step. The resulting data can subsequently be processed with common neural network architectures.
  2. Using specially designed network models capable to handle irregular data sources, like point clouds, directly within their architectural concept - often at the cost of performance though.

In his master's thesis, which Paul wrote at senswork, he implemented a method that is capable of directly accepting point clouds as input data to train a neural network. This neural network enables 3D objects to be localized within a larger point cloud regardless of their position, orientation, or scaling.

This tool will soon be part of our deep learning software "Neuralyze". Stay tuned!

 

Go back

It will load necessary cookies, Google Fonts, Google Maps, OpenStreetMap, Youtube and Google Analytics. More details in our privacy policy and our imprint.