Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

wrong results for large images #7

Open
leoneckert opened this issue Jun 20, 2016 · 8 comments
Open

wrong results for large images #7

leoneckert opened this issue Jun 20, 2016 · 8 comments

Comments

@leoneckert
Copy link

Hello, do you have an idea why I am getting so very different results for the same image depending on its dimensions? If it is too big the program doesn't seem to be doing the right thing after detecting the facial landmarks.
Any help would be very much appreciated, I'll attach screenshots of two versions of the same image.
Thanks!
Here the large version:
screenshot from 2016-06-20 14 53 11 copy
And the small one:
screenshot from 2016-06-20 14 54 49 copy

@dougsouza
Copy link
Owner

Hi @leoneckert,

Yes, this is known issue, and was already discussed here #5. We still don't have a fix for that rather than resizing the input to a smaller dimension.

Cheers.

Doug

@leoneckert
Copy link
Author

Okay, thanks for coming back to me! I thought maybe a solution has been found already.

Is the problem maybe the 3D model that's being used? Feels like the model has those particular dimensions and trying to "wrap" something larger around it, results get crazy... Just a feeling, have been digging into the code, but couldn't rhyme together any other explanation to the problem :)

Anyhow, thanks again

@mwlodarczyk
Copy link

Hi @dougsouza,

Any update on the problem with large images? Any clue where we can search for a bug when debugging?

Best,
Michal

@dougsouza
Copy link
Owner

@mwlodarczyk,

To be honest, I haven't worked on this project lately. Regarding the issue: I think the problem may be related to size of the 3D model and the reference image that are used to calibrate the camera. I think that part is a good place to start debugging.

Cheers,

Doug

@TalHassner
Copy link
Collaborator

@mwlodarczyk,

I'm also uncertain why this happens. There really shouldn't be any reason for this, especially if you're not using symmetry (which is the only part I believe depends on the size of the input image). I may be wrong and it has been some time since I looked at the code.

I wonder, however, why you don't simply rescale the input image to a size that works, as a simple solution.

If this leads to problems with landmark detection (i.e., you need the high resolution to detect the landmarks) you can simply detect the landmarks on the high rez image, then rescale it AND the landmark coordinates.

Did you try something like this?

Tal

@mwlodarczyk
Copy link

@TalHassner,

Once I resized the image, the code works fine. However, I want to try using this face frontalization for biometrics recognition purposes. Therefore, I want to have high quality image to get the best possible results. When resizing, the obtained image quality is not good enough for my purposes.

Currently, I am busy with some other issues. I will get back to this code around the middle of September. I will let you know if I find anything interesting.

Best,
Michal

@oxydron
Copy link

oxydron commented Oct 4, 2016

Anyone know which dimension should a face have to work okay?

Best regards

@oxydron
Copy link

oxydron commented Oct 4, 2016

Some tests suggests that if the face height belongs to [150,220] interval pixels, the algorithm works just fine. While we no fix is made, I suggest to resize faces proportionally with height in this interval.

Best regards

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

5 participants