search

TadasBaltrusaitis / OpenFace

OpenFace – a state-of-the art tool intended for facial landmark detection, head pose estimation, facial action unit recognition, and eye-gaze estimation.
Other
6.95k stars 1.85k forks source link

Greyscale image output #967

Open TimChaouoft opened 3 years ago

TimChaouoft commented 3 years ago

Describe the bug First of all, I would like the thank the developer for the wonderful program! I have 300 greyscale images of headshots. I wanted to get their head pose information. When I ran ./bin/FaceLandmarkImg -f, 285 of the 300 pictures generated output, which is great!

However, there were 15 images that did not generate any output. Even if I tried to change the model/algorithm .txt, I still got 285 outputs (i.e., csv). Has anyone encountered similar problem?

Here is an example of my 4 targets, one of them (i.e., R70) did not produce any output:

Screen Shot 2021-05-10 at 9 22 10 PM

One way I tried to solve it is that I ran ./bin/FeatureExtraction -fdir with these face images, including the unreadable R70. Although I know this command is designed for videos..When I ran this, all outputs were generated as frames. However, the confidence for the unreadable R70 (frame 2 in this case) remained low (0.47), the in "success" it says 0. Please see the screenshot of my feature extraction.

Screen Shot 2021-05-10 at 8 47 40 PM

I wonder if I could still use the information generated by this output?

Is there anyway to get the rest of the 15 images' output using either the facelandmarklmg or feature extraction? Is there a way to use a different landmark detection algorithm so that these images could be processed?

I'm a student at UofT doing a project. I would really appreciate it if I can get a solution! Thank you very much! [tim.chao@mail.utoronto.ca]

Desktop (please complete the following information):

Additional context Example images. R70 is the one that didn't produce any output if I run Facelandmarklmg R70 R69 R71 R72

TimChaouoft commented 3 years ago

R70 This image is R70 and couldn't be processed.

TimChaouoft commented 3 years ago

Is there a way to use a different landmark detection algorithm so that these images could be processed? Thank you!

TadasBaltrusaitis commented 3 years ago

However, there were 15 images that did not generate any output.

I don't think this is related to images being grayscale, but I think it is more ot do with face detection not having worked on those images. One thing you can try is changing the face detector used internally by OpenFace, but this would require changing code and rebuilding (the change would be fairly minor though)