Séminaire IPI : « Industrial Part Pose Estimation from Virtual Images with Deep Neural Networks » par Julien Langlois
26 avril @ 14 h 00 min - 15 h 00 min
Le prochain séminaire IPI aura lieu
vendredi 26 avril 2019 de 14h à 15h, en salle D005 à Polytech.
Julien LANGLOIS, doctorant au sein de l’équipe IPI parlera de : « Industrial Part Pose Estimation from Virtual Images with Deep Neural Networks« .
Résumé : Automation of the bin picking task still remains challenging when the parts are dark and poorly textured. The complexity is even higher when only classical RGB images can be used in an industrial environment where the light is not controlled (parts might be glossy and show strong light saliencies). We present a neural network technique using a single top-view 2D image of a bin to segment several parts within the image as patches. An encoder-decoder network is first inferring the depth map from the patches
then, two other networks give the object orientation and Z-translation in the scene. An ICP based pose refinement is employed using the retro-projected depth map and the object CAD model. As the training dataset creation might be time consuming and difficult to obtain (the part needs to be maintained on an unbalanced pose), all the networks are
trained using synthetic images. With several light parameters and material reflectivities, the proposed pipeline can absorb the introduced virtual-real bias.
La présentation sera disponible sur le cloud : https://uncloud.univ-nantes.fr/index.php/s/xcLeXnAEs6JgC5p
The next IPI seminar will held on Friday
the 26th of April (2pm-3pm).
The room (at polytech) will be D005.
The speaker will be Julien Langlois who is PhD student in the IPI team.
See title & abstract above.