“It is not possible to separate between plastic and paper”

We promised different than what we will deliver. In the presentation it should be clear that the separation will be made between objects and these objects will be classified as metal, plastic and paper.

Even if we did this, the classes were not thought beforehand. This is a very important part of the project and requirements, because it may inviabilizate the project.

The network model, dataset creation and its training were also a big uncertainty problem brought by Heitor. He suggested a transfer learning model to extract image features and then a network to detect the objects from those features.

Action plan

• Decide, for each material, all objects that will be recognized
• Do a proof of concept for the training process
  • Web scrapping to get images from the objects and generate dataset
  • Create network with infra as suggested by Heitor (or another one that solves the problem)
  • Check the precision of the network and if it is viable (e.g. It could work with larger dataset)

Deliverables

• [x] List of objects that will be classified, considering the most common trashes in universities

  • Main responsible: Ayrton
  • Classes chosen
    • cardboard
    • crumpled_paper
    • juice_box
    • metal_can
    • paper_bag
    • paper_ball
    • paper_cup
    • plastic_bag
    • plastic_bottle
    • plastic_cup
    • plastic_cutlery
    • snack_bag

• [x] Report with dataset specification (e.g. how to label, images requirements, needs boundary box or not, how many images per object, etc.

  • Main responsible: Ian

• [x] List of objects that will be used for the proof of concept, considering the classes that are most common and most critical (similar to each other)

  • Main responsible: Ayrton
    • crumpled_paper
    • metal_can
    • paper_ball
    • plastic_bag
    • plastic_bottle
    • paper_bag

• [x] Dataset creation for proof of concept

  • Main responsible: Ayrton
  • The dataset image files can be found here:

  Classes - Google Drive

• [ ] Report with transfer learning model and how it will be applied in our case. Specify EVERYTHING (extractor, language, augmentation methods, layers removed, etc. as if it were a paper)

  • Main responsible: Ian

• [ ] Report with the results of the proof of concept, with adequate performance measurement considering the classes and the network used

  • Main responsible: Ian

• [x] Risk response plan for case of proof of concept goes wrong

  • Main responsible: Waine

  Check risks management

“Metal sensors are not precise enough, and should detect anywhere”

The project, as presented, suggests that the object can be anywhere in the chamber and the sensor should detect it from anywhere. This implies also that the object could be anywhere for the image.

Actions

• Specify and present better the space limitation of the object
• Use image processing to detect metals as well