Can semidense_observations be used as keypoint ground-truth for training a feature extraction model?

[banafshebamdad]

Hello,

I am working on a feature extraction model and I am considering using semidense_observations as keypoint ground-truth (ground-truth for feature locations) for training. Since these observations correspond to tracked 2D points, I wonder if they are reliable enough for feature learning?

Best regards,
Banafshe