Landmark Data For custom use

[Abeltemi]

I truly appreciate your great effort in putting this out.

please, how or where do i get the landmarks, such in the format of demo_landmarks.json that is required to be passed to the model.
Also is it possible to have a following pipeline: Takes 2D image for instances extract the landmarks then run the estimate_measurements.py to get the body measurement.

[Abeltemi]

@DavidBoja Please kindly help out. Thank you

[DavidBoja]

Hi,

we used the ground truth landmarks from the CAESAR dataset in the paper.
To the best of my knowledge, there is currently no open-source implementation for finding the landmarks on 3D scans.
If you find one, please do let me know.

If you don't want to develop your own landmark detector, you might circumvent that by fitting an SMPL using SMPL-Fitting and then transferring the landmarks defined on the SMPL to the scan - you can follow the annotation process here.
The fitting usually requires landmarks for initialization and grounding the optimization, but if your scan and template are already close, maybe you could do it. Otherwise, you could circumvent using the landmarks for the optimization by using a marker annotator, something like this one. Find the markers on both the SMPL and scan, use SMPL-Fitting to fit the SMPL onto the scan using the markers (along with other loss terms), and then transfer the landmarks from the SMPL to the scan.

For the question related to 2D the answer is maybe. You can't know until you try. You can try and train a landmark annotator, lift the landmarks in 3D and then run the method. Note that your landmarks need to be in the A-pose.