RESULTS

[rafaloz]

eval_metrics_geom_model.json

[rafaloz]

[ruaridhg]

So geometry augmentations are more important than brightness/contrast augmentations

[mxochicale]

Many thanks @rafaloz and @ruaridhg for your contributions, I will add plots in our slides.

• It would be great if you can share a google-colabs notebook or the notebook that you use to create the plots.
• Can you please link the methods for geometric and brightness/contrast augmentations?
• Do you mean ground truth in the bar results for original?

Thanks, Miguel

[rafaloz]

Hey Miguel,

1. The code to make the plot is shared here:

It is a very simple plot, I included the results from the .json files copying and pasting:

https://drive.google.com/file/d/1gxnIe41Z21t2I2UlX3i1NL6_fR90IRak/view?usp=sharing

2. The geometric augmentation included:

   1. crop
   2. random rotations
   3. Elastic transformation
   4. Vertical flip
   5. Horizontal flip

3. The code can be found at this link: https://drive.google.com/file/d/10VlnozEQqWJ0va6J14xMW7opyLZviZWv/view?usp=sharing.

4. The original label refers to the model trained for 100 epochs with all augmentations, which I believe is the ground truth.

Rafa

[ruaridhg]

Hey Miguel

The augmentation I changed was RandomBrightnessContrast which adjusts both the brightness and contrast and the code can be found https://colab.research.google.com/drive/1CgdpeDY4SLpKddk6dFvXpFwcx0skF8dj?usp=sharing

I created the .json files in this notebook https://colab.research.google.com/drive/1j67zkMhi6PS0l9IZR5mjS33jwhpAGL0v?usp=sharing