Model Zoo Hand Posture and Motion Detection

[ganbolede]

Hello,

I am trying to implement a gesture recognition application.

I am using 53L8A1 to collect data.

And found out the model zoo supports Hand Posture using 2D CNN. In the model zoo is uses minimum distance, max_distance and background distance. Based on the normalization it follows:
X1_norm[zheaders["distance_mm"]] = (X1[zheaders["distance_mm"]] - 295) / 196
X1_norm[zheaders["signal_per_spad"]] = (X1[zheaders["signal_per_spad"]] - 281) / 452

I would like to know how number 295, 196 for distance and 281, 452 is calculated.
Since i am trying to detect dynamic gesture as well with motion.
It possible to do it in 3DCNN with 8x8x2xNumber Of Frames. Where 2 are the distances and the signal per spad?

[liaifat85]

With 8x8x2xNumber of Frames, you can represent an 8x8 grid (reflecting spatial resolution in each frame) across two channels (distance and signal per SPAD) over a time series of frames (Number of Frames).

[labussiy · Accepted solution]

Hello,

Actually, these values have been calculated on the public ST dataset, they are used to normalize the data between -1 and 1.

In a perfect world, we should update it for each new dataset.

Another approach could be to use the min/max distance as the normalization parameters.

Regarding your question on the model to detect dynamic gesture, I will be very honest, I never tried to use a 3D CNN. Why not trying a RNN (Recurrent Neural Network)? 

Also, if you are planning to deploy your model on STM32, you should check which topology is supported by the STM32Cube.AI.