What is the difference between online and offline?

[wzr0108]

[jtattershall-ga]

Offline uses all the data, because it has it available Online takes in a window_len worth of frames; window_len//2 old and window_len//2 new fames per inference. It therefore has an inherent delay associated with the model - which is not ideal if robot policies rely on up-to-date keypoint

[wzr0108]

For offline videos, which model provides more accurate predictions?