Running the pose model with empty frame when low confidence detector output.

[arturoptophys]

Addresses issue #137 with a quick-and-dirty trick: when low-confidence prediction from detector -> pass a tensor of zeros to the pose model instead of raw frame. This way this frame will not have valid predictions, but the DLCLive will not crash and continue with next frame.

[sneakers-the-rat]

i can't really evaluate this, since i don't know how the pytorch model works, but it looks reasonable - doesn't seem like it should be possible to fall back to an array with a totally different shape dependent on quality of estimation