“Deep Learning no limits”
On 17th October 2018 at Buildo, Data Science Milan has organized an event about 3D image processing. Deep learning on 2D images has been achieved good results on classification tasks thanks to the use of Convolutional Neural Networks and the availability of data. Now, 3D data are growing in fast way.
“3D Point Cloud Analysis using Deep Learning”, by SK Reddy, Chief Product Officer AI in Hexagon
In this talk were showed several technologies used to manage 3D point clouds, so what is the mean of point cloud?
Point cloud is a database containing points in the three-dimensional coordinate system. It is a very accurate digital record of an object or space and it is saved in form of a large amount of points that cover surfaces of an identified object.
Tasks with point cloud can be shared into neural network challenges (unstructured grid data for CNN filters, invariance to permutations of point clouds, the number of points changes depending from the sensor used) and data challenges (the use of scanned models bring missing data, noise from sensors used and rotation implies different point clouds).
Octree-based Convolutional Neural Network (CNN) for 3D shape analysis (O-CNN) is built upon the octree representation of 3D shapes. It takes for input the average of vectors from 3D model sampled and performs 3D CNN operations on the octants occupied by the 3D profile surface. O-CNN supports numerous CNN architectures and works for 3D images in different representations. Look out the github repository.
The architecture approach of PointNet is the use of a single symmetric function: max pooling. The network learns a set of optimization functions/criteria that select informative points of the point cloud and encode the purpose for their selection. The last fully connected layers of the network aggregate these learnt optimal values into the global descriptor for the entire shape (shape classification) or they are used to predict point labels (shape segmentation). Check out the code
SPLATNet is based on architecture that process point clouds without any pre-processing, it takes point clouds as input and computes hierarchical and spatially features with lattice filters; it allows easy mapping of 2D information into 3D images and vice-versa. Apply the code
Check out the video of the event.
Author: Claudio Giancaterino
Actuary & Data Science Enthusiast