Size of neural network input
livingbeams opened this issue · 1 comments
livingbeams commented
Congratulations on the great work.
This is more a question than an issue.
In the comment in InputStream.cs is said that the network input size is set to:
const int Width = 640 + 1;
const int Height = 352 + 1;
Because 16 * n + 1 dimensions are required.
May I ask where do these constraint come from?
I've tried other arbitrary sizes such as 1280x720 or 1920x1080 and it seems to work properly.
Regards.
keijiro commented
It came from BodyPix: https://github.com/tensorflow/tfjs-models/blob/master/body-pix/src/util.ts#L72
You can safely use 1281x721 or 1281x1073, but 1280x720 and 1920x1080 may introduce degradation. In my experience, it introduces small offset in output.