Back to all Internships

Thumb

Internship

Seraing

Euresys sa

3D object registration with deep learning

Company description


Euresys est un leader dans le domaine de l’acquisition et du traitement de l’image, avec plus de 25 ans d’expertise dans les domaines de la vision industrielle et de la vidéo surveillance. Avec de grands clients internationaux et des technologies uniques, Euresys est actif dans l’acquisition vidéo analogique et digital, la compression d’image, le control de caméra, l’analyse et le transport de l’image. Nos produits combinent des logiciels performants et les dernières technologies de la microélectronique.

Euresys est situé sur le Liège Science Park.

 

Job description

Context


3D object registration consists in finding the 3D pose of an object in a point cloud based on a 3D model (CAD file) or a reference scan (point cloud).  This technique can be used in belt-picking (Figure 1) and bin-picking (Figure 2).

 

Figure 1 Bin-picking, https://www.photonics.com/Articles/Advancements_in_3D_Vision_Software_Bolster_Bin/a65477

Figure 2 Belt-picking, https://www.engineerlive.com/content/new-robot-guidance-part-localisation

Existing techniques are slow and not very robust. Deep learning methods could outperform them.
A solution could be to train a neural network to predict the 3D pose of an object:

  • From a CAD file, generate synthetic 2.5D image (ZMap) in a multitude of orientations.
  • The input of the network would be a ZMap (Grayscale image) and the output would be the 3D pose.
  • A possible network architecture could be similar to YOLO (an object detector) but instead of predicting offsets of a bounding box, it would predict offsets of a 3D pose.
  • This network would only work with the specific objects it has been trained on (no generalization compared to object detection).

Objectives

The intern that will join the Euresys Vision Software team will have to:

  • Compare state-of-the-art algorithms for 3D object registration and similar problems.
  • Choose/create the most suitable network architecture for the problem.
  • Find/create a dataset and a metric for the problem (partially done already).
  • Train the network and evaluate the experimental results obtained on the dataset.

This internship will take place in the Vision Software team under the supervision of an experienced software engineer.

Searched Profile

Requirements

The intern should have knowledge of python, computer vision and deep learning. Experience with C++, pytorch, tensorflow, GPU programming and 3D data processing is a plus.

The internship will have a minimum duration of 10 weeks.

We offer

Une expérience dans une équipe de recherche et développement: travail en groupe, méthodes de développement, outils de gestion de projet...

Le sujet est à la frontière de la recherche académique et du développement industriel.

Le stagiaire sera encadré par un ingénieur de R&D. 

Apply for this job
Only for undergraduate students

Articles you may want to read