Getting wrong landamrks with each config file
Nibenitez opened this issue · 7 comments
Hi!
Thank you for sharing your great work with the community.
I would like to ask wich config file should I use for the model below, I tried geometry and depth, but without sucess. Is it something wrong with my mesh? Thank you.
Hi,
I am happy that you use the tool.
It could be a problem with the initial orientation and placement of your scan. Try to take a look at the section in the README file on: "Specifying a pre-transformation"
You can experiment with different angles and see if your results get better.
Best,
Rasmus
Thanks Rasmus for your prompt reply and guidance!
I tried different angles but still without better results. I checked in Meshlab the orientation and I have the mesh over the z and x axis.
I will keep trying, but I would appreciate any other recomendation, maybe I missing something else.
Could you send the scan to me? Then I can see what might be the problem.
Could you send the scan to me? Then I can see what might be the problem.
Thank you! I have many of them to detect landmarks and your tool will be very usefull!
I made it work by changing the scale=100 in the pre-align section of the JSON config file.
It is currently assumed that the face are recorded in physical space (in millimeters). For non-calibrated 3D scanners, it is (unfortunately) necessary to scale the scan so it is approximately the same size as a real face (in mm).
Wow! Thank you Rasmus! I am getting great results now by changing the scale=100 per your comment.
Another last question, I am trying to evaluate meshes predicted by different facial reconstruction methods versus my own scans, and to obtein any metric about the quality of the reconstruction. I am using the following code: https://github.com/patrikhuber/fg2018-competition. In this code, is a requirement to have 7 landmarks:
Is it possible to change the output of the .txt file with those only 7 landmarks instead of 84? So, I can use it in the evaluation algorithm. It is just a question, I know that probably the model was trained to detect more landmarks. Otherwise I can code a python script to get a .txt with only the 7 key points from the 84 txt file.
Thank you again for your time and for sharing your great work!
I am happy that it works. I think it would be the easiest that you create a separate script that filters the landmarks from 84 to 7.