Information about the calibration procedure mentioned in the paper

[gabrielecaddeo]

Hi all,
Thanks for sharing with us your great work.

I read the paper and in the appendix you mentioned a necessary calibration step to be performed (I attach here the screen).

However, I could not find any part of the code where you actually perform such calibration. I also saw that your renderer is designed to load a different background image given an id from the user, however this seems to require the following method from the tacto class that I could not find in the associated python package.

MidasTouch/midastouch/render/digit_renderer.py

Line 49 in 8e67453

background=cv2.imread(tacto.get_background_image_path(bg_id)),

Can I have more information?

Thanks,
Gabriele

[suddhu]

Hi Gabiele! Thank you for reaching out.

• The calibration step here is generating the training data for the TDN in TACTO, but using an assortment of background images. You can think about it as a form of data augmentation to generalize to a wider range of DIGIT sensors. I currently haven't added the data generation scripts for the TDN, but I can work on getting that open-sourced as soon as possible. Would that be of interest to you?
• get_background_image_path is a function in our modified tacto repository that the conda env should automatically install. You can find it referenced here.

[gabrielecaddeo]

Thanks for the fast reply!

Would that be of interest to you?

Yes, for sure!

You can find it referenced here.

Thanks for the clarification

[suddhu]

Hi @gabrielecaddeo! The data generation scripts are now live in midastouch/data_gen from the recent commit 13f0887. Let me know if you have any further questions, and thank you for expressing interest in it!

I'll close this issue for now, but can re-open if you would like to follow-up.