CopernNet : Point Cloud Segmentation using ActiveSampling Transformers

Published on: June 24, 2024

A broad interest, a passion for coding, and a healthy dose of curiosity made Louis Regout choose a position at Kapernikov. Today, Louis is mostly working on projects relating to asset data, AI on point clouds and AI development for the Telraam traffic counting project.

Louis started at Kapernikov early 2021. Before that, he obtained a Master’s degree in Applied Mathematics. But Louis’ interests also extend to other areas: next to obtaining minor degrees in law and entrepreneurship, he learned to program through self-study and even developed a few personal projects, including a 3D rendering engine and a file-sharing application.

With his scientific approach and mathematical background, Louis decided to step off the beaten path and design a new architecture for point cloud segmentation. He published a scientific paper to show how his solution worked faster and better than the old solutions that were mainly based on “feature-engineering”.

Revolutionizing Railway Maintenance: Meet CopernNet

In the dynamic field of railway maintenance, accurate data is critical. From ensuring the health of vegetation surrounding the tracks to maintaining the integrity of rails and poles, precise 3D data is invaluable. That’s where CopernNet comes in—a cutting-edge tool designed to transform how we handle and process this vital data.

The Heart of CopernNet: Advanced Technology

CopernNet isn’t just another data processing tool; it’s built on the powerful “Transformer architecture,” previously renowned for its capabilities in language translation and image recognition. Adapting this technology for 3D spatial data, CopernNet offers a new level of insight into the complex geometries of railway environments.

Learning Like a Pro

CopernNet starts its learning process by absorbing the geometric patterns within vast volumes of point cloud data, akin to a student who first tries to understand the overarching themes of a textbook without focusing on the minutiae. This phase allows CopernNet to grasp the general structure and layout of railways, including tracks, catenaries, and surrounding vegetation. As its training progresses, it then moves to more detailed, labeled examples, honing its ability to recognize and interpret specific features with precision.

A Closer Look

Initially, CopernNet scans the railway environment in a grid-like pattern, sampling data points broadly to get a general sense of the space. As it identifies areas of interest—perhaps a section of track that shows signs of wear or a densely vegetated area threatening to encroach on the line—it shifts its focus. This targeted approach allows CopernNet to concentrate on the most relevant parts of the data, refining its understanding and ensuring that critical issues are not overlooked.

Speed and Precision

When it comes to speed, CopernNet truly shines. It can process 28,000 points per second, translating to about 250 meters of track per minute. Imagine measuring every centimeter of a cable with such precision that it feels like a specialist has examined each millimeter in person—this is the level of detail and efficiency CopernNet brings to railway maintenance.

Beating Traditional Methods

Traditional methods of processing railway data involve a challenging step known as “feature engineering” or “manual feature selection.” This process requires experts to meticulously identify and define the specific characteristics—like the spacing of sleepers or the wear on the rail—that the system should focus on for analysis. This approach not only demands considerable effort and expertise but also leads to bottlenecks in quality, unpredictable project timelines, and solutions that are tailored to specific problems without easy adaptability. Additionally, experts skilled in feature selection are scarce, which can slow down development. CopernNet revolutionizes this by automating the selection of these characteristics, dramatically reducing the need for expert input and enhancing both the speed and accuracy of data processing.

Why Does This Matter?

For industries reliant on railway infrastructure, CopernNet is not just a technological advancement; it’s a paradigm shift. By making 3D data processing faster, more accurate, and more accessible, CopernNet enables railway companies to perform more frequent and detailed inspections without disrupting service. This leads to safer, more reliable railway operations and opens new possibilities for maintaining and improving critical infrastructure.

In essence, CopernNet is transforming the railway industry by providing the tools to see and understand their operations like never before, paving the way for innovation and efficiency in maintenance and planning.


We asked Louis some questions on his design :
Q : “What led you to do scientific research in this area ?”

L : “Well, first of all, I find this an exciting and active area of scientific research ! I also thought that publishing my design might lead to improvement suggestions from academia and offer opportunities to learn. As part of Kapernikov, we are happy that we could contribute to the revolution that AI brings. I also secretly hope that this helps to convince interested candidates who would like to join the Kapernikov ecosystem and do exciting and fascinating work!

Q : “Why did you decide to write a paper on this topic ?”

L : “The old solutions were mainly based on “feature-engineering”. A feature is some measurement in an area. For example : How high is this point ? How flat is this region in space ? But this is a lot of work and it is a quality limiting factor. Also it is unpredictable regarding development time, problem specific and last but not least, it is super hard to hire feature engineers. The previous scientific research on this topic proved to be bad for very big point clouds, such as the ones that Kapernikov works with. I wanted to design something fast, that could work with very big point clouds (+10 billions points in total). In essence, I wanted to design something simple because very complex solutions are hard to maintain.

Q : “What are the results of the CopernNet invention ?”

L : “CopernNet allows us to add new functionalities to our system much, much faster. It’s just a matter of training data. Additionally, we will deliver much higher quality, e.g. we handle every corner case better. On top, we can be less expensive with this kind of solution. Which is good news for our customers !”

Louis submitted this paper for NeurIPS 2024. 1

Kapernikov is very proud of Louis’ achievement ; he is part of a team that strives for excellence in a cutting-edge field.

  1. The Thirty-Eighth Annual Conference on Neural Information Processing Systems (NeurIPS 2024) is an interdisciplinary conference that brings together researchers in machine learning, neuroscience, statistics, optimization, computer vision, natural language processing, life sciences, natural sciences, social sciences, and other adjacent fields. They invite submissions presenting new and original research on these topics every year. ↩︎