make-sense
make-sense copied to clipboard
SkalskiP
Pre-Annotation Data Cleaning
Is your feature request related to a problem? Please describe.
Before the process of annotation, I believe many people go through the process of data cleaning: things like image-level labels (brightness, image quality,...). Furthermore, there might be a need to crop the images with a certain aspect ratio with a certain scale for the image.
Describe the solution you'd like Having a labeling and cropping tool for data cleaning with a batch ofdata.
Hi @ssakhavi I understand that in this situation, when we download labels, we will also have to download a whole package of photos? Because the training should take place on new (edited) photos?
By right, the training should be done on the edited photos, yes.
@SkalskiP But maybe we don't need to download the whole images. We will need to download the process that has led to the editted image and then we can generate the new images and labels via the downloaded image.
As an example: Imagine we need to crop image A. We don't need to dowload the cropped image A but only the 4-points that lead to the new frame.
@ssakhavi I am sceptical about such solutions for (at least) two reasons:
- This approach is "sufficient" if we are only talking about crop, but what about rotation? What about CV operations like filters, binarization or other potentially useful operations?
- I want MakeSense to be as intuitive as possible. Exporting a photo is obvious for the user. Nobody has to guess what kind of information was returned from the application.... I uploaded the photo, cropped it, exported it and got the cropped photo.
If we were to go into such a thing, it would have to be a separate module, maybe with a separate tab. Where I can do it all. And then export photos and labels.
Yes. Exporting the image is the most obvious way. You're right.
On Fri, 4 Oct 2019, 17:30 Piotr Skalski, [email protected] wrote:
@ssakhavi https://github.com/ssakhavi I am sceptical about such solutions for (at least) two reasons:
- This approach is "sufficient" if we are only talking about crop, but what about rotation? What about CV operations like filters, binarization or other potentially useful operations?
- I want MakeSense to be as intuitive as possible. Exporting a photo is obvious for the user. Nobody has to guess what kind of information was returned from the application.... I uploaded the photo, cropped it, exported it and got the cropped photo.
If we were to go into such a thing, it would have to be a separate module, maybe with a separate tab. Where I can do it all. And then export photos and labels.
— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/SkalskiP/make-sense/issues/73?email_source=notifications&email_token=ABOAG63WDWR2WQB3J5DGGXDQM4EMLA5CNFSM4I5FROIKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEALC6UA#issuecomment-538324816, or mute the thread https://github.com/notifications/unsubscribe-auth/ABOAG6ZENZCLPRQCLJCU6MLQM4EMLANCNFSM4I5FROIA .