search

facebookresearch / supervision-by-registration

Supervision-by-Registration: An Unsupervised Approach to Improve the Precision of Facial Landmark Detectors
Other
762 stars 165 forks source link

The effect of landmark detection dramatically decreases as the distance increases #59

Open xfj81525 opened 5 years ago

xfj81525 commented 5 years ago

Hi, as the distance between person and the camera increases, the landmark detection effect dramatically decreases, does anyone have the same problem .

D-X-Y commented 5 years ago

Does distance increase during training or evaluation?

xfj81525 commented 5 years ago

Hi D-X-Y, firstly i use camera to captre face and detect landmark on live, i am close to the camera (within one meter), face landmark detection works well and stable. then i use a video file to test landmark detection, in which the person may be 4 meters away from the camera, the detection effects decreased dramatically, even the the landmark points were not complete i try to use the pretrained model(https://download.pytorch.org/models/vgg16-397923af.pth) to do the same testing , but unfortunately it can't be loaded ,the error is it failed at param = snapshot['args'], reporting KeyError: 'args'.

so i wonder whether it is because of the algorithm or my traininng procedure that make this phenomenon. (my training data contains 10 thounds images) thanks in advance.

xfj81525 commented 5 years ago

Does somebody have the same issue as I describle, thanks