Reduce variation in output?

[bradleybauer]

Is it possible to reduce the difference between the output of the network for two different but very similar inputs?

I'm asking because if I run py webcam.py ... and sit still in front of the webcam, then the output of the network is very different from frame to frame.

[eridgd]

This effect is often referred to as 'flickering' and is common to nearly all neural style transfer techniques, as far as I'm aware. I've seen a few papers on video style transfer that address this with variations on a 'temporal consistency' loss that tracks pixel correspondence across frames using optical flow. Some examples:

• https://blogs.nvidia.com/blog/2016/05/25/deep-learning-paints-videos/ and for code https:/manuelruder/artistic-videos. This gives great results but takes minutes to process a single frame due to the optimization process.

• https://arxiv.org/pdf/1705.02092.pdf - This should run in real-time but the code has yet to be released, see Not an issue - Characterizing and Improving Stability in Neural Style Transfer jcjohnson/fast-neural-style#132

• https:/rtqichen/style-swap - Produces consistent stylizations without optical flow, but the overall results are less impressive than other methods.

[bradleybauer]

Thanks for the info.