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

Doubts about Video Tags #48

Closed zrqi closed 5 years ago

zrqi commented 5 years ago

Hello, author, thank you for contributing the code, I have a question about the video label in the SBR process. According to what you said in the paper, the video can be unlabeled, but in the SBR process, the frame t-1 should be labeld? I still have the other question. Since each frame requires the position information of the previous frame, is it necessary to give a label to each frame during training? Because each frame is the previous frame of another frame.

D-X-Y commented 5 years ago

Thanks for your interest. The frame t-1 is not necessary to be labeled. Each frame requires the position information of the previous frame, which is provided by the learned detector instead of the manual label.

zrqi commented 5 years ago

In the detection loss, you need to calculate the loss between the annotation and the detection results. Isn't this calculated for frame t-1?

D-X-Y commented 5 years ago

You can calculate the detection loss for any labeled data (not necessary to be frame t-1). We use frame t-1 as an example to visualize our high-level idea in a simple way.

zrqi commented 5 years ago

Oh! I see，thank you