FOSS4G 2022 general tracks

On the road to 3D semantic segmentation
2022-08-24, 14:15–14:45 (Europe/Rome), General online

A few years after the democratization of semantic segmentation advanced techniques in the 2D
context (imagery), there are more and more initiatives for exploiting such algorithms with 3D
datasets. The context appears favorable: public and private initiatives are arising in terms of
massive 3D dataset collection, hence a huge amount of 3D point cloud data will become available in
a near future. As an example, the french cartography institute (IGN) is currently targetting a
full-coverage of the country with LIDAR data in the five next years.

Considering 3D point clouds is a really challenging task regarding semantic segmentation. Whilst
this data format allows to represent a scene with a high level of details, the unordered and
unstructured nature of the data makes the standard convolution neural network approach
ineffective. However other deep learning algorithms exist to cope with these
characteristics. Depending on the desired accuracy and the labelled data availability, some
"softer" machine learning approaches may also complete the toolbox.

Leveraging georeferenced data in such a context may be an interesting avenue in order to improve
the algorithm performances. In any case, these fairly innovative solutions can be applied in some
geographical use cases, e.g. cartography with street-views, Building Information Modelling (BIM),
...

This presentation will provide some insights on these 3D semantic segmentation related topics:

  • the 3D semantic segmentation state-of-the-art will be flied over;

  • the BIM use case will be detailed through the presentation of an ongoing R&D project carried out
    by Bimadata.io, Oslandia and the LIRIS lab (CNRS);

  • the geo3dfeatures project (https://gitlab.com/Oslandia/geo3dfeatures) will be showcased in
    order to illustrate what the seminal component of a 3D point cloud segmentation software could
    be.

Data engineer at Oslandia for 5 years, I'm working with data on a daily basis. I spend my time on data modelling, data analysis, geospatial data pipeline conception problems. I do some AI too (2D- and 3D semantic segmentation), when Oslandia has the opportunity to work on R&D topics!