This topic contains 1 reply, has 2 voices, and was last updated by 
pers_d7pyza March 1, 2020 at 12:26 am.
-
Posts
-
pers_d7pyza
KeymasterPrediction module Level 1 – Looks at a) what app is running, b) recent history of trigger events – e.g. widget presses/mouse movements, c) recent AV output – makes prediction for t+1,t+2,…t+N AV output
Prediction model Level 2 – Looks at the output of Level 1, the uncompressed output, along with the user inputs + the inputs about latency/lag, etc. – tries to use utility theory to optimize a cost function – low bandwidth is good, lag, & sharp jitter are bad, short term blur is medium bad,…. what would be the optimum outputs for the predictive codec considering the cost function?
Level 3 – is the actual software implementation of the strategy from Level 2.
Algorithms that get too complex can reach a point of diminishing returns. But conceptually, the optimization is not just picking best prediction vs. using the regular codec. Consider the case of a mouse event that triggers a rotation display of a graphical object. For some cost functions, the optimal insta-strategy might be to start sending a hierarchical/wavelet coding of the upcoming prediction while waiting a few samples to see if it is looks correct rather than sending the uncompressed output immediately. Why? If the prediction looks wrong, then you don’t jitter – you change course with some unnoticeable lag, while if the prediction is correct then you catch up with some unnoticeable lang & low bandwidth usage for compression.
You must be logged in to reply to this topic.