Présentation Iris de Gélis (IRISA) : Change detection and characterization from 3D point clouds using deep learning methods.

Iris de Gélis, PhD Student from IRISA will come to present her work to SDC and SERTIT on the Change detection and characterization from 3D point clouds using deep learning methods.

If you are interested in such work, please feel free to attend.

When: 12th October, 9h30

Summary :

To perform a map update or to identify affected areas in case of disasters, the automatic detection and characterization of changes is an important step. In a number of environments such as urban areas, most changes occur along the vertical axis (e.g. construction or demolition of buildings) and the use of 3D data such as point clouds is therefore interesting. However, 3D point clouds are very particular data because, unlike images, they are not distributed on regular grids but they are sparse and unordered data. To cope with these specificities, a number of methods perform a prior discretization of the data into Digital Surface Models (DSM) in order to apply usual image processing tools. This is not optimal because there may be a significant loss of information. Other approaches analyze point clouds directly. This last family has not yet been approached from a deep learning perspective. Thus I will present a method for change detection and characterization based on a Siamese network using convolution kernels operating on 3D points directly. The method has been designed on the basis of the good results obtained by the Siamese architectures in change detection in 2D images and by the KPConv network in semantic segmentation of 3D point clouds.