Thursday, September 12th, 2019 (Discussion Critical Images with Silvan, Labeling App)¶
Axel, Ulf, Silvan, Subir, Michael, Malin, Maren, Katha, Josefine, Luana
Discussion points of today:¶
label app - some decisions still open
discuss critical images - Silvan’s opinion
tasks for the next weeks
Presentation of the lableapp:¶
shows all hree images of one asparagus, the pictures are preprocessed
labels go into a csv file
select a folder/datei to save the data
then open labeling dialog to hand label assistant
aswer the qustions asked with yes/no
display additional information
peak in the back to of the figure as indicator to detect violet
Katha and Malin are going to write a manual for the labeling process
aim: doublelable some of the images
we still have to decide on how we handle intra personal differences in labeling
discussion lapeling app:¶
possible to show only the violett colour chanel - check pixels in the violet range
take out violet question
take out thickness questions
double labelling with label app for “labeled folders” first - before we do other stuff
possibility to zoom?
float for some 0 and 1?
general remarks concerning sorting by Silvan¶
sort 1A anna more conservatively
aim to have more than 50% of the first class in the end
minimal violett —> violett , even if we detect a tiny bit of violet, it already counts as violet
minimal rust -> not rusty, but still first class, only if rust is at head or really bad, it counts as rust
discussion in general / for deep learning:¶
data augmentation : mirroring maybe not so useful (as miror right image might be very similar to left image)
maybe normalise the data —> network can be simpler
rotation of pieces
training different nets (one with background one without)
we want to learn features, and not final labels
from manual labelling we get impression what features are important, what not! —> use this for nets , we can better “understand”
in one of the next sessions ulf can present sth about neural nets
also: for technical questions we can always come to his office
TASKS:¶
finish labeling app
label images
continue manual feature extraction
for next step: research on why to use different approaches for deep learning nets - not TOO many approaches
each who does net training: looks around what model & technique to use
decision: do we want to use tensorflow?
rz logins to Ulf
NEXT STEP/ SUGGESTIONS:¶
maybe in future also make it possible that machine can detect class only from one picture - maybe machine is not able to always take 3 pictures - and then can’t classify with only one!?!