Training result shows discontiunity in prediction?

[gurbee]

Thank you for the open source of event to depth network! Currently I am training an event to depth model based on the model structure mentioned in your paper, however the result of the predictions are very discontinuous, which predictions varies a lot from this frame to the next one, as demonstrated as the follow gif. My model utilize all the settings in your paper, including model structure, loss(including multi-scale output), training settings etc. I would like to know have you encounter this kind of problem(discontinuity), and what is the solution of it, thank you for your work, looking forward for your reply.

Thank you for your help, looking forward for your reply!!

https://imgur.com/a/cBFeFHw