Advice

• Brave image search api can help you find images. You add kind of a cue and it searches for you.

Pose Model

Very good for interactive visualization. Is the most popular one to use for these artistic visualizations.

Pose Estimation

• Detect human figures
• Estimates the key points and joints
• Nose detection is pretty stable

Done with PoseNet (2017). Also used in teachable machine. Is a Convolutional Neural Network. They trained the model with examples. They used keypoint heatmap and the offset vectors to calculate the xy position of different keypoints.

Input: RGB Input

Output: Positions (check article of tensor flow

If the model has more points in the grid to check from it will be slower but have more accuracy. Related to that. So if less points, faster and less accuracy. The model is pretty fast anyway and we don’t have a latency issue anymore. This with mobileNet

AFter that, there will be feature maps related to how many points in the grid are. They will calculate the x,y position based in the highest probability to calculate where is that point.

To find more 3D human pose they use GHUM model to detect the keypoint. They collected real world data. But because there’s not enough data they also synthesized more in 3D engines with their 3D models and key points. Lots of data sets for this models are a mix of real and synthesized data.

We’ll use ml5.bodyPose

moveNet vs BlazePose: BlazePose is more accurate because it measures more points.

BlazePose has 33 keypoints and is fast enough to track fast movement. Has more points in wrists face and ancles. Also more stable!