Getting confidence values from the body points about tf2-mobile-2d-single-pose-estimation

[AngryDuckFTW]

I was wondering the same, I used the iOS version which has a confidence value in the multi array output but the android one seems to have a "Probability" I assume thats the same. Set probability and get probability are never called but the set probability function has commented out code that puts a float value in the heatmap array at 0, index. Which unless im been stupid would be an array so how is setting a float there? and how can I get It out. CoreML model is doing something similar but I can't see how this java implementation would even work

[tucan9389]

@JakeWade5 @AngryDuckFTW

The deep learning models(Core ML model and TFLite model both) take the multi-dimensional vector and return the multi-dimensional vector.
The output vector of pose estimation is usually heatmaps(or confidence map).
In our case, there are 14 heatmaps as output vector.
Each heatmap has the location information of each human part(like left elbow, right shoulder and nose etc.).
On a heatmap, you can get the maximum element's value and the column, row index.

The maximum element's value will be confidence value.
And maximum element's column, row index will be the position of the part.

If the maximum element's value is sooooo low, you can filter with a threshold.

[soans1994]

@tucan9389

how can we train coco model for multipserson heatmap prediction?
I we train the whole image without cropping with single person label for each sample, does the model learn to predict multiple persons?

thank you