Szabolcs Lőrincz is a 23-year old scientist from Transylvania. He recently defended his master’s thesis at Aalto University, which he completed working at the Finnish Geospatial Research Institute FGI, Autonomous Mapping and Driving research group. You can watch his video interview below, where he tells what benefits his project brings to the scientific community and to the society, or read below his interview.

Text: Szabolcs Lőrincz & Annukka Pekkarinen. Video: Annukka Pekkarinen/FGI, 

Tell us, who are you?

I am Szabolcs Lőrincz, 23 years old, from Transylvania. I’m currently in the final stages of being a Master’s student at Aalto University and of being a research scientist at FGI.

And what is your research about?

My research topic is about automatically classifying millions of points in three dimensional point clouds captured by LIDAR sensors in urban traffic environments. The classes of interest are roads, sidewalks, road markings, traffic lights and signs, cars or pedestrians. We aim to tackle two major issues present in current methods aiming to solve this problem:

First, some classes appear with different frequency than others in these environments, known as class imbalance. This holds especially for three dimensional scans of urban traffic environments, since the large surfaces such as roads and sidewalks are with orders of magnitude more frequent than the small objects such as traffic signs.

Another issue is that large scale annotated data sets are needed for training deep learning models, the annotation process being burdensome and inefficient in terms of time and human resource.

To this end, we experimented with several class-imbalance handling techniques additionally to adopting data efficient methods from the image processing literature to the task of point cloud segmentation and showed that each design choice boosted model performance.

What is the greater benefit your research may create?

As mentioned before, most methods, even data efficient ones, usually need to be trained on manually annotated data sets to tackle specific tasks.
Therefore, I manually annotated a large scale data set captured by FGI’s Roamer in order to be able to properly test our methods. The result of this annotation is the second largest data set in terms of number of points annotated in urban traffic scenarios. Additionally, our data set contains scans from two different LIDAR sensors operating on different wavelengths, being a distinctive feature compared to current publicly available data sets. This can propel the wider research community towards developing and testing new methods leveraging multi-spectral information in point cloud processing research.

Additionally, our method takes us one step closer towards data efficient point cloud segmentation methods, being one of the most important aspects in terms of reducing human labour needed for manual annotation.

For the greater public, these point cloud processing methods can be applied in autonomous driving, both in real-time for localization and navigation, and in mobile mapping, usable for instance for virtual tourism or disaster management.

What’s next?

I was always fascinated by conducting research and that’s why FGI was a great fit, I’m thankful for making it possible for me to experience working in a truly research-focused community. For now, I am moving home to Transylvania and since I couldn’t find any suitable PhD position back home, for now I am looking for a remote job in Finland. I would still like to work in research and development, possibly in autonomous driving or medical fields. Later on, a PhD is definitely on the horizon.