Friday, June 14, 2019 (Planning of last days of data collection, Subgroup divisions)

Participants: Luana, Maren, Malin, Sophia, Richard, Josefine, Axel, Ulf

Updates

Labelled Data

How many (labelled) pictures do we need?

  • we don’t always have all asparagus classes (not evenly distributed)

  • only one week left

  • having more data won’t hurt

  • what fraction of data might be labelled? (prob. less than 10%)

  • semi supervised learning techniques might recommend how much labelled data we need

Schedule for trips to Rheine to collect labelled data

  • Monday 17.06 : Richard, Josefine

  • Tuesday 18.06 : Sophia, Luana

  • Wednesday 19.06 : Richard, Malin

  • Thursday 20.06 : Maren, Luana

  • Friday 21.06 : Sophia, Josefine

  • still open: 14., 15., 16. & 22., 23.

Problems and Thoughts

Ideas of bucket-wise collection

  • don’t think about perfect classification piecewise but bucketwise

  • find solution for problem, that even sorted asparagus is sorted by human eye and not perfect as the machine would need to learn the classes

  • use a system where you update while classifying

  • apply this idea to features and not images

After data collection

  • have subgroups with smaller tasks

  • try all sorts of possibilities for sorting (ANN etc.)

Curvature extraction

  • actual asparagus sometimes S-shaped, bump at end etc.

  • method should be able to detect different kinds of curvature (return all asparagus that is not totally straight as curved)

User interface design

  • how should the user interact and also influence the sorting-process?

  • if people here are interested in this, we include it (but it is no must)

  • maybe difficult to say right now while we do not have an algorithm or anything yet

  • decision postponed to later date

Subgroups

Decision app

  • Michael

  • gives some information on the shown asparagus picture

  • has only yes and no buttons

Documentation

  • Katha

  • (gather more literature)

Feature extraction

  • flower, color and maybe rust still problematic

Framework

  • already start it and then implement it as soon as feature extraction is done

Literature research

  • rather a task distributed to all subgroups

Organisatorial stuff

  • don’t have too many subgroups

  • distribute groups and the assign more than one task to them

Report

  • already start the report on our study project

  • not only write what we have done but also why we tried it (make this very clear why we thought this is a good idea), how we arrived there and also put in what failed

  • generally good: before you start something, write down what you expect/plan to gain from it and always update what worked and what didn’t

  • around 100 pages expected, 10 pages per person (no final number)

Next week

Presentations

  • data augmentation

  • semi-supervised learning

Superresolution