Simple question. Does YOLOv7 pose estimation model output 3D points? ... And, how to get them?
NOTE: I have already run the
test.py over custom images (successfully) with the object detection model.
I am also seeking 3D pose estimation tool too.
Sadly, output of YOLOv7 is 2D, the x, y and the confidence score, so there is a 3*17=51 length tensor.
There is a guy tried to use 2 cameras and YOLO to do the 3D estimation, you may take a look. https://github.com/SkalskiP/sport
As far as I know, YOLOv7 is for 2D pose estimation for multi-person pose-estimation, where models like MediaPipe does single-person pose-estimation.
For the 3D pose estimation, I am using the "3D-MPPE" model, since the pretrained models are provided. It is a single person 3D Pose Estimation model.
The 3D-MPPE model has 2 inner models: RootNet and PoseNet.
The RootNet estimates the root depth, which is the z-axis value of the target (relative distance from camera to the target). And the 3D-MPPE PoseNet estimates both x and y axis values of each target joints of the person.
Below are links for 3D-MPPE official repos (for both PoseNet and RootNet), and article for actual usage of the 3D-MPPE model: