Pose estimation with OpenIMAJ

[yosaba]

Hi,

As part of a head tracking application, I need to do pose estimation to determine where the person is looking. Is there a standard way to do that in OpanIMAJ?

I have seen suggestions for using the cv.solvePnPRansac method in OpenCV (https://docs.opencv.org/master/d7/d53/tutorial_py_pose.html). I know there are a number of standalone pose estimation projects, e.g., OpenFace, OpenPose and DeepCutLab. I have seen examples of loading an OpenPose model into OpenCV and processing an image with it using an OpenCV DNN (https://gist.github.com/berak/eef67a1afb4b6445d117ea1a0135c8db).

How would you do similar things in OpenIMAJ?

Thanks

[jonhare]

Don't know if it will work well enough for your application, but the CLMFaceDetector can give an estimated roll, pitch and yaw of a head; of course it doesn't deal with where the eyes are looking though....

[yosaba]

Thanks jonhare! This looks great -- I should be able to get those from the CLMDetectedFace objects it finds.

If it turns out I need to use something like a model generated by OpenPose, is there a straightforward way to do that?

[yosaba]

I implemented a test using CLMFaceDetector and fortunately happened to scroll back through all the debug on the console and found that it uses code governed by a strictly non-commercial license. My project is under a research grant but the desire is to make it a product at some point so it looks like CLMFaceDetector won't be an option.

Are there other ways to get pose orientation using OpenIMAJ, either within its code base or using something like OpenPose or OpenFace?

Thanks!