Prototype 1 and next steps
Published in
Jan 18, 2021
As demonstrated, we are currently able to reliably identify the skeleton of images. 45 out of 50 images from 10 different genres were correctly processed. The incorrect ones are mostly accurate enough for our purposes. Some examples are shown below:
The OpenPose package has a potential to perform 3D and multi-person matching, which may be explored in the future.
Each picture roughly takes 5 seconds to calculate, and the performance is limited by hardware. We think the latency is acceptable for this version of our game and might improve it when necessary.
Next, we will start to work on matching as separated in two stages:
- Compare the skeleton to ensure the posture is correct.
- Use motion detection to verify the displacement.
Skeleton comparison is more difficult, as we are currently not aware of direct methods or packages to use.
