Is there a video Demo to show it's performance?

[justusschock]

Is there a video Demo to show it's performance?

[visonpon]

actually, i'm interest in your model's performance VS dlib or openpose.

[xjcvip007]

@justusschock ,maybe 300vw datatset would meets your needs link: https://ibug.doc.ic.ac.uk/resources/300-VW/, or you can follow this demo (https://wywu.github.io/projects/LAB/LAB.html).

[justusschock]

Not yet, but all results are given in the paper. I will have a look whether I can provide a video. In what kind of performance are you interested? Speed or accuracy? Because the first thing can't be shown by a video at all

[visonpon]

Accuracy.
Hope we can see the video performance soon,especially about human face. thanks~

[justusschock]

Do you have a good idea on how to provide the video? I don't want to add it to version control due to it's size.

[justusschock]

Yeah, but I don't know, how to provide it

[justusschock]

@visonpon on which data do you want me to show the performance? I could use some random images, but that would not cover the tracking capability...

[xjcvip007]

@justusschock , thanks for your amazing job~ One question, you mentioned that the method is a super real-time method in your paper, so why not test on a facial video?

[justusschock]

@xjcvip007 I'd like to do so, but I'm not aware of any publicly available dataset containing facial videos. Do you know one?

[justusschock]

Update: I got it working on a video from 300-VW...

I'm just evaluating, which format would be best to display this, and will upload it later on

[justusschock]

Demonstration Videos comparing our method to dlib can be found here as overlay and here as side-by-side view

Sorry, I can't embed them directly (they are too large).

Evaluation Method:

• initialization:
  • Detect Faces
  • Predict Shapes based on Initial Bounding Box
• tracking:
  • Predict Shapes based on Bounding Box extracted from shapenet's prediction of the previous image

My Conclusion: Both methods have similar accuracy, ours is a bit more robust (updating the bounding box from dlib's predictions led to a collapse), recovery from small errors is possible, huge different in speed (our method is far faster).