M2P 2023

Recomposition of Parts Into Objects in Industrial Applications

  • Roffilli, Matteo (Bioretics srl)
  • Castelli, Filippo (Bioretics srl)
  • Flebus, Emanuele (Bioretics srl)
  • Neri, Mattia (Bioretics srl)

Please login to view abstract download link

When people look at the world, objects naturally emerge in their minds as a whole. Machines - and the algorithms built into them - often work differently. They proceed in a bottom-up fashion by first recognizing the parts (or sub-parts such as textures) and then recomposing all the information to a higher level of knowledge, thus reaching the whole object. In this scenario, we present a parameter-free theoretical approach to clearly define where an object is located. The technique acts at the heatmap level of knowledge, an intermediate representation where each pixel/voxel of the original data is assigned with a probability of pertaining to a specific target object. Some results borrowed from industrial applications (automated grading and sorting of fresh fruit) are presented to demonstrate the viability of this technique in demanding industrial environments.