Probability/score of a keypoint

[marcin-adamczewski]

Hi,

Thanks for this sample :) I'm wondering how to get the probability/score of each keypoint in Android example? The output of tlLite is heatMapArray, which contains only positions of each keypoint.

Btw, is there any documentation how input and output for a specific model should look like?

[edvardHua]

The heatMapArray shape is [1, 96, 96, 14] which means [batch_size, height, width, channel].

Channel num here is the keypoints indices.

Therefore the maximum of each [96, 96] is the probability of a keypoints.

I will update the documentation later.

[marcin-adamczewski]

Thank you @edvardHua. I was able to extract those values.

If I may ask, why do you apply Gaussian filter on the heatmaps values? Should I use values after filtration or before to find the probability? I've noticed that the values are bit smaller after filtration.