Noise Removal from Point Clouds & Meshes


Website Mainblades

Mainblades’ drone uses a high-end lidar to reconstruct 3D-models of commercial aircraft – see the
top image. We use these models is various forms in our product, a completely automated visual
inspection of an aircraft.

Our 3D models are remarkably accurate – we can estimate the dimensions of any aircraft up to
centimeter accuracy with respect to manual-specifications. However, sometimes we miss some data
in a point cloud – or the surface is simply a bit noisy while we know aircraft hulls are smooth. In the
process of making these 3D models there are various sources of error: the robot localization, noise
on the actual measurement or artifacts that are inherent to the used data-structure of the 3D model
(e.g. raw point clouds, voxel-maps, meshes).

In this internship we want to experiment with techniques that improve our models: in particular
mitigating holes and smoothening surfaces. Ideally we would end up with a few post-processing
steps that can improve some aspects of our models for specific use-cases. As intern you will
implement model processing scripts that do this (preferably in Python) using libraries like Open3d.

Relevant reading material:


About Mainblades

Mainblades is an innovative company located in The Hague which develops automated aircraft
inspections, automating the whole process from visual drone inspection up to report-generation. We
are looking for a student whose curiosity is triggered by this internship. Ideally the candidate is a
master-student with some programming experience, familiar with Linux and a general hands-on
mentality. The duration of the internship can be 3 to 4 months and the starting date is flexible. This
is a paid internship with a monthly compensation of €500.
In order to apply, please supply:

  • A recent CV with relevant experience
  •  A short description about yourself
  •  A little assignment: find a 3D model on the internet and save it as:
    • a pointcloud
    • a mesh
    • a voxel map

note that for this assignment, we advice to use some open source software (e.g. CloudCompare) and
limit your own time. When you can not find a random model, resort to the classic ones: Attach the files in a compressed folder and
make sure it’s not too big (e.g. < 5 mb).

Thomas Horstink

To apply for this job email your details to