Easy-Label-Ordinality-Regression (ELO-Regression)

A simple GUI python script for labeling relative ordinality between images

Why labeling expert/subjective regression problem is hard?

1. Only a domain expert is able to give direct score. (e.g. Which one has better muscle? Can you score them?)
2. Subjective label depends on a single person. (e.g. Which one do I prefer?)
3. One cannot give a globally consistent score between [0,100] for large-scale images.

Why ordinality can help?

1. Average people can tell which one is better even without the absolute score.
2. Very few choices. (Left is better. Right is better. I cannot tell. Bad pair.)

How do I run this thing?

# Download your images set into one single folder (e.g. './download').
# Run command
python ELO.py


# 1. Selected the image folder you want to label.
# 2. Labeling!!!
# 3. Save the label into npz file.

Labeling command

• Left (←): Left is better. Save as number 0
• Down (↓): I cannot tell which one is better. Save as number 1
• Right (→): Right is better. Save as number 2
• Up (↑): Bad Pair. Save as number -1

System command

• q: Quit the program.
• b: Back to previous pair.
• s: Save the current results.

Dependencies

• python
• tkinter
• numpy
• PIL
• glob

To do

• Resume the label results

Contact

If you find this respository useful, or you want to do some cooperation.

Feel free to contact me: Tsun-Yi Yang, shamangary@hotmail.com